It won’t work:

My suspicion was that GPT 5.6-Cyber would succeed, but the frequency and manner of its success removed all doubt. We have to reassess sandboxing quality for capable AI agents, and in general the software stack with which they interact.

An off-the-shelf VM is not enough to contain a modern, cyber-capable AI agent. There is simply too much attack surface. Even innocuous features (like running with a display) add extra, exploitable attack surface.

This is interesting:

The records describe a force-generation mechanism for several General Staff components, including the GRU, Main Operational Directorate, and 8th Directorate, which is associated with protected communications, cryptography, and information security.

[…]

The reporting also linked a 2024 Department No. 4 graduate, Aleksei Kondrashov, to Military Unit 74455, widely known as Sandworm.

That unit has been associated with destructive cyber activity against Ukraine and other targets, including the 2017 NotPetya attack.

The reports do not establish that every listed graduate participated in a named operation; assignments should therefore be described as reported unit placements, not proof of individual operational involvement.

The Bauman material reframes Russia’s cyber capability as an institutional system, not merely a collection of well-known threat groups.

It suggests that Moscow has formalized a recurring pathway from university recruitment to military service, where students receive supervised technical and ideological preparation before entering intelligence, cyber, and security roles.

For defenders, the leak reinforces the need to track Russian operations as a combined threat: espionage, destructive activity, military reconnaissance, technical surveillance, and influence campaigns may draw on related personnel pipelines and overlapping doctrine.

The exposure of Department No. 4 also provides researchers with a clearer lens for understanding how the GRU sustains cyber capacity beyond the familiar APT28 and Sandworm brand names.

Authorities in Australia have arrested two men believed to be members of TeamPCP, a prolific cybercrime and data extortion group blamed for perpetrating the longest running spree of software supply chain attacks ever.

In a statement released today, the Australian Federal Police (AFP) said two men from Western Australia, aged 21 and 23, were arrested in connection with a “sophisticated cybercrime syndicate that allegedly created malicious open-source software to rob thousands of global businesses.”

The AFP did not name the defendants, but KrebsOnSecurity learned the 21-year-old suspect’s real identity in June, and has been communicating with him ever since. This story includes interviews with TeamPCP’s self-described spokesperson, and examines clues left behind by the TeamPCP leader that likely led to his undoing.

TeamPCP vaulted onto the cybercrime scene in late 2025, embedding malicious code in hundreds of open source software tools and extorting victims for profit. Members of the group made headlines by compromising corporate cloud environments using a self-propagating worm dubbed Shai-Hulud, which added malicious code to open source programs maintained by developers whose credentials at public code repositories like GitHub or NPM were phished or stolen.

Writing for Wired, journalist Andy Greenberg described TeamPCP’s core tactic as a kind of cyclical exploitation of software developers.

“The hackers gain access to a network where an open source tool commonly used by coders is being developed,” Greenberg wrote in May. “The hackers plant malware in the tool that ends up on other software developers’ machines, including some who are writing other tools intended to be used by coders. The malware allows TeamPCP’s hackers to steal credentials that let them publish malicious versions of those software development tools, too. The cycle repeats, and TeamPCP’s collection of breached networks grows.”

TeamPCP also has practiced something akin to cyclical recruitment. In May, the source code for the third iteration of Shai-Hulud was published online, and TeamPCP soon after launched a contest offering $1,000 in virtual currency to whichever participant could conduct the largest supply chain operation using the worm’s code. According to the contest rules, participants were scored based on the number of weekly and monthly downloads of packages they compromised — directly incentivizing them to target the most popular code libraries.

A screenshot of a message from TeamPCP’s Telegram account, announcing the supply chain hacking contest. Image: dataminr.com.

“TeamPCP has stated the competition is a recruiting opportunity and they intend to purchase all meaningful access harvested from participants’ campaigns,” the security firm Dataminr wrote. “The $1,000 XMR (Monero) prize is a recruitment floor and has been dismissed by the actor as ‘just like participation trophy,’ adding ‘if you find something good you will be paid way more,’ confirming the contest’s true function as talent identification and malicious access acquisition at scale.”

In March, TeamPCP executed a supply chain attack targeting AI infrastructure by compromising the code for LiteLLM, an open source AI gateway that connects users to more than 100 different large language models. A recent analysis by the security firm CloudSEK found TeamPCPs attack on LiteLLM harvested cloud service keys and other secrets from more than 2,500 organizations, including many of the world’s top technology companies.

In May, TeamPCP claimed credit for compromising at least 3,800 code repositories at the Microsoft-owned GitHub, after a GitHub developer installed a code extension that was compromised by TeamPCP’s malware.

MEET THE CYBERCATS

Security experts say TeamPCP is less of a hacker group than an amalgamation of threat actors from multiple cybercriminal gangs who sometimes work together toward similar goals.

“It is not a structured criminal crew with a single operator,” said Austin Larsen, a principal threat analyst with the Google Threat Intelligence Group. “It is a peer community of individually-skilled actors, with one clear center of gravity.”

That center of gravity is George Prepakis, an accomplished security researcher and self-described exploit developer who operates the Twitter/X profile @kernelstub. Earlier this year, @kernelstub tweeted a public invite link to a Matrix chat server he created and dubbed “Cybercats,” and TeamPCP and several other cybercrime entities have been using this server to communicate daily for the past several months.

A screenshot of the Matrix chat server “Cybercats,” whose members used hacker handles associated with multiple distinct cybercrime groups that have occasionally collaborated on a series of supply chain and data ransom attacks over the past nine months.

Kernelstub, like other administrators in the Cybercats chat, has been using his Twitter/X profile name as his handle in these Matrix communications, frequently tweeting references to other members and to conversations taking place in the Cybercats chat. In a number of cases, the corresponding X accounts for members of the Cybercats chat taunted cybercrime victims publicly before the incidents were reported in the news media.

The Cybercats administrator listed at the top of the screenshot above — “Boxturtle” — is a close associate of TeamPCP who has been tweeting about the group’s conquests under the name @xpl0itrsturtle. This handle corresponds to a data breach broker active on Breachforums and Darkforums who has been selling data stolen in a wave of recent breaches at automobile manufacturers, including BMW Group, Audi, Honda, Mercedes-Benz, Volvo and Toyota, as well as data allegedly taken from Snapchat and SportRadar.

The data leak site for the extortion group or handle “xpl0itrs.”

The Cybercats administrator “SeesawSec” in the screenshot above is the alias of whoever is behind the cybercrime group known as Fulcrumsec, which recently claimed credit for data extortion attacks against the pharmaceutical giant Novo Nordisk, the data broker LexisNexis, and Avnet, a Fortune 500 distributor of electronic components.

The data leak site of Fulcrum Security, a.k.a. Fulcrumsec.

The Cybercats administrator “@pcpcasper” also has been using a similar name on X to discuss TeamPCP’s attacks and victims. This person has an extensive message history on Telegram, where their messages and shared videos show @pcpcasper is an active and vocal member of the National Socialist Network, a neo-Nazi political organization based in Australia.

At one point in these chats, @pcpcasper shared videos and images of what they claimed was their cat, and several of those videos place this user in Western Australia. One source close to the investigation told KrebsOnSecurity that @pcpcasper was one of the two arrested, a claim supported by messages that @kernelstub posted online this morning.

The Cybercats member roster pictured above also features an administrator with the username “T,” which is short for the now-banned Twitter/X profile @pcpcats, the account operated by the self-described TeamPCP spokesperson who was arrested today. As we’ll see in a moment, @pcpcats also is from Western Australia.

By the time @kernelstub tweeted a public invite link to the Cybercats Matrix server, T/@pcpcats was posting only infrequently to the group chat, with other members often inquiring as to his whereabouts and well-being. The group’s collective concern related to @pcpcats’s tendency to blame his increasingly extended absences on the use of hallucinogens and other narcotics that kept him awake for days on end, but also caused him to crash in bed for several days after the highs wore off.

WHO IS THE TEAMPCP LEADER?

The Cybercats member @pcpcats has used multiple nicknames on the cybercrime forums, including EllisD25/LSD on Darkforums, BulkDMT on Breachstars, and Express on Breachforums. These accounts are linked because they all advertised the same Tox ID and/or Session ID as instant message contact handles in their cybercrime forum posts. BulkDMT was also known on the forums as DMT Host, which was a virtual private server (VPS) hosting service that was peddled on Darkforums and Breachstars.

DMT Host/EllisD25, posting on the English-language cybercrime community DarkForums in September 2025. Image: ke-la.com.

According to the cyber intelligence firm Intel 471, Express registered on Breachforums using the email address shitstickpp@gmail.com. Intel 471 finds Express posted on Breachforums across a two-month period in 2025 using four different Internet addresses located in South Africa. On July 30, 2025, Express announced on Breachforums they were selling access to 14 gigabytes of data stolen from South Africa’s State Information Technology Agency.

The threat intelligence platform Flashpoint recorded more than a year’s worth of messages from the TeamPCP leader’s alter ego on Telegram — Persy_PCP —  who claimed they split their life living between two countries [full disclosure: Flashpoint is an advertiser on this blog]. “I have these [files] as well, problem is these are in another country,” Persy_PCP explained to another user inquiring about a stolen data set in November 2025.

Later that month, Persy_PCP complained, “My whole country is racist and they want people like me dead.” Flashpoint records show BulkDMT shared in September 2025 that “this country is going to fucking starve when they take the farmers land,” a likely reference to white landowners in South Africa who claim to be targeted by an ongoing genocide campaign.

This tracks with public reporting on TeamPCP. Cyberscoop reported in June that Google had traced TeamPCP’s residential and mobile Internet address connections to South Africa, “indicating the primary operator was located there during at least some of its attacks.”

BulkDMT also shared on the group chat at Breachforums that they were recovering from an addiction to methamphetamine. “My life is kinda fucked rn [right now], but that’s fine and there isn’t really a point in pouring so much emotional energy into that fact, my parents had money but I unfortunately got really addicted to some things so I don’t get to benefit from that. As long as I continue to survive, stay sober, and move closer towards my goals that’s enough drive and meaning.”

The identity threat protection company SpyCloud finds shitstickpp@gmail.com shows up in the registration of an account called ChristmasSnow on the cybercrime community Raidforums in 2022. Nearly all of the Internet addresses used to access that account came from ISPs in Perth, Australia, SpyCloud found.

KrebsOnSecurity looked up all of those Perth IP addresses in passive DNS records maintained by DomainTools.com, and found one of them — 211.27.196.111 — for several years was used as a private file server by a family in Perth with the last name of Thomson. Those records show at least three hosts — ithomson.direct.quickconnect.to (a remote Synology server), kthomson0061.direct.quickconnect.to, and joshuawthomson39.myqnapcloud.com (a QNAP network storage device) — persisted at that address between 2022 and 2025.

Searching on “joshuathomson39” in the breach tracking service Constella Intelligence reveals an account at the freight forwarding company kwe.com created in the name of Joshua Thomson from Perth, Australia. The open source intelligence platform Epieos finds the phone number attached to that kwe.com account was used to register a Facebook profile for Josh Thomson, which says his family includes a brother named Ruben, his father Ian, and his mom Cindy.

That Facebook profile also says Josh and his family are originally from Pietermaritzburg, in KwaZulu-Natal, South Africa, but currently living in Cottesloe, a beach-side suburb of Perth. A search in DomainTools for Ian Thomson and Australia unearthed five domains by the same registrant, including securecomputing.au, thomson.org.au, and thomsonfamily.net.au. Ian Thomson is a dentist in Cottesloe, and a biography says he graduated from The University of the Witwatersrand in Johannesburg, South Africa.

Constella finds a joshua@thomson.org.au registered a number of accounts online, but Josh doesn’t seem to have much of a connection to dodgy cybercrime forums. His brother Ruben, on the other hand, has quite the presence on these communities, dating back to at least 2018. Constella reports ruben@thomson.org.au frequently reused the password “joshuathomson1,” and Constella further finds that password was used by just a handful of accounts, including yolosolo17@gmail.com and surfinup8@gmail.com.

According to Intel 471, surfinup8@gmail.com was used to register the user Yolosolo17 on the crime forum Altenen in 2018, and that user account was registered from the Perth address 110.141.230.15. On Altenen, Yolosolo17 advertised free web proxies, as well as the domain rubenthomson.com, which was at one point used to sell steeply discounted iPhones. DomainTools says rubenthomson.com was hosted at 110.141.230.15 and registered to surfinup8@gmail.com.

A cached copy of the domain rubenthomson.com from 2017 shows a login page underneath a banded stack of money. Image: archive.org.

SpyCloud reports 10.141.230.15 was used by the email address sheepstealing@gmail.com on Raidforums and surfinup8@gmail.com on Nulled, and that the same IP was used by the email addresses ian@thomsonfamily.net.au, jasper@yakuza.cc, and rubenthomson1@gmail.com. SpyCloud also shows that sheepstealing Gmail address is tied to the accounts Sheep420, YoloSolo117 and Yakuza.cc on Raidforums, and to the account “Sheep Stealing” on Hackforums. Intel 471 says sheepstealing@gmail.com was used to register the account DingoFlour on Breachforums in October 2023, as well Sheepx on Altenen.

Epieos reports that ruben@securecomputing.au is tied to an Airbnb account for Ruben, who described himself as a Web developer who went to school at the University of Western Australia and was living outside the country. “Hey, I’m Ruben, my friends call me Ellis. I’m a Perth creative who occasionally books rooms when visiting family and for photography.”

Epieos also finds sheepstealing@gmail.com registered an upwork.com profile under the name Ruben, who said his main skills are setting up secure server hosting solutions and PHP full-stack Web development.

“I’m familiar with Linux, working with relational databases (SQL),” the Upwork profile reads. “I also script in Python mainly for writing social media bots.”

The Upwork profile for Ruben Thomson in Cottesloe, Australia.

Epieos further discovered sheepstealing@gmail.com is connected to a Microsoft account for Ruben Thomson, and to a now-defunct GitHub account called XmasSnow/XmasSnowisBack that scammed people on the forums in 2022 by claiming to sell exclusive exploits for recently-released software patches (recall that shitstickpp@gmail.com was used to register a forum account named ChristmasSnow).

This same sheepstealing email address registered a Twitter/X account in 2026 called “Gone Fishing” that lists its location as South Africa. That Gmail account also left several reviews for businesses listed on Google Maps over the past seven years, but all of those establishments are located on the west coast of Australia.

Business reviews in Western Australia left by the Google account sheepstealing at gmail.com.

The people search service Pipl finds a 21-year-old Ruben Thomson in Western Australia who has a phone number ending in 979. A lookup on that number at Epieos reveals it is connected to a TikTok account under the name Ellis, and to a PayPal account in the name of Ruben Thomson.

Finally, a search on the name Ruben Thomson from Cottesloe at the Australian government’s record of registered businesses finds he has incorporated or served as an official in multiple companies created since 2024, including Secure Computing Solutions, Tensor Industries, and another entity ironically named OPSEC Express. Recall that Express was BulkDMT’s nickname on Breachforums.

Australian companies connected to Ruben Thomson. Image: abr.business.gov.au.

It’s ironic because OPSEC is short for the term “operational security,” which refers to techniques and behaviors used to obfuscate and compartmentalize one’s real-life identity online, and using your cybercrime handle as part of your own company name is very much the antithesis of that practice.

There is at least one other major opsec failure by Ruben that exposed a link to TeamPCP. In June 2025, someone using the name Ruben Thomson registered on HackerOne, a popular “bug bounty” program that seeks to reward and recognize researchers who agree to work with affected software vendors to help fix the flaws before publishing about their findings. What was Ruben Thomson’s chosen HackerOne username? Deadcatx3, a nickname that has been flagged by multiple security firms as an alias used by TeamPCP.

The HackerOne profile for “Ruben Thomson” uses the nickname Deadcatx3, which multiple security firms have concluded is an alias used by TeamPCP. Image credit: flare.io.

INTERVIEW WITH ELLIS

In early July 2026, not long after having discovered clues about Ellis’s real life identity, KrebsOnSecurity interviewed the TeamPCP leader via Signal, where he was remarkably open about his activities and personal struggles [for the sake of simplicity, the TeamPCP spokesperson will be referred to from here on as Ellis].

Ellis claims he stopped doing cybercrime for TeamPCP in March 2026 — just before the attacks that compromised LiteLLM — and that at least one other individual has taken over the group’s leadership since then. Ellis shared that a year earlier he had just completed the latest in a series of detox and sobriety programs, and was two months sober when he reconnected with some old friends from the malware development scene.

“One year ago I needed help monetizing some [GitHub credentials], I was two months sober and needed a distraction and something to keep busy as well as people to speak to,” Ellis said. “I had largely disconnected from my old circle, they had become very toxic and I needed to get away from the substances. Previously I had done some mass exploitation campaigns and grew up doing [malware development] and [capture the flag] contests. There were some friends who were also vending but had stopped a while, and one of them introduced me to some chats where I posted access for sale.”

Prior to that, Ellis said, he was homeless and hopping between “some very unstable places.”

“Blackhatting is fun,” he said. “There are actual rewards and incentives to learn and you grow with your team. Without qualifications, no employer will even take the time to hear you out.”

Ellis claims he’s earned a grand total of about $20,000 for his activities with TeamPCP, and that it was never about the money or fame for him. Asked whether his experiences with TeamPCP might prepare him for gainful employment in a legitimate IT job, Ellis said he doubted it.

“I am nowhere close to a skill level where I am comfortable, and this would take maybe half a decade of further experience,” he said. “I no longer have to choose between rent and food for that I’m grateful and so are the team members.”

Ellis expressed no remorse over his cybercrime activities, and said he was grateful for the friendships and relationships built throughout his engagement with TeamPCP. The young hacker also seemed resigned to his fate, and told KrebsOnSecurity that he’ll accept the consequences if he’s ever arrested.

“If I’ve already been found out then its out of my control, I’ll make peace with that,” he said. “Honestly, I think someone like me needs a lot of help that prison just can’t provide. If I had the funds to study different parts of the field and closer guidance, this would have turned out differently. But that’s a pipe dream and we both know this.”

It is clear from reading Ellis’s posts to the group’s Matrix server chats that his struggles with sobriety are ongoing. On Thursday, June 25, Ellis told @kernelstub he was about to “trip” with his “homie.”

“What kind,” @kernelstub inquired.

“Ketty and some DMT,” Ellis replied, referring to the dissociative anesthetic ketamine and dimethyltryptamine (DMT), a powerful psychedelic compound that is found naturally in some plants but is also synthetically produced in underground lab environments. “There’s a little 2cb so we might throw that in the mix,” he continued, referring to another psychedelic compound by its chemical shorthand.

Roughly two weeks before his arrest, Ellis told KrebsOnSecurity he was ready to leave his life of crime behind and was prepared to turn himself in, but that in the meantime he was making plans to tie up loose ends.

Less than 24 hours later, the TeamPCP leader posted an image on Telegram showing a yellowish powdered substance in a baggie and on a scale, possibly synthetic DMT. The image shows the powder being weighed next to a series of small vape cartridges, two of which are open on the table in front of the photographer.

An image posted by the TeamPCP leader to Telegram, advertising his acquisition of some type of psychoactive substance, most likely a synthetic version of the powerful hallucinogen known as DMT.

The two defendants were arrested Wednesday morning. The AFP said the men face a combined 14 cybercrime offenses and are scheduled to appear in Perth Magistrates Court today.

Charlie Eriksen is a security researcher at Aikido Security who has closely followed TeamPCP’s cybercrime campaigns. Eriksen said TeamPCP are a good example of a new kind of threat actor that does not fit neatly into the usual categories.

“They are not a state actor, not quite organized cybercrime, and not purely ideological,” he said. “Their motivations seem to mix money, disruption, attention, and ideology.”

Eriksen said that historically there has always been a meaningful gap between reading about an attack technique and being able to reliably turn it into an operational campaign, but that large language models (LLMs) and artificial intelligence increasingly are helping threat actors to bypass that knowledge gap.

“You had to understand the research, adapt the code, troubleshoot it, build infrastructure around it, and then repeat that process across different targets,” he said. “LLMs have compressed that gap significantly.”

According to Eriksen, this creates an environment where threat actors suddenly have the ability to operate at significant scale without having developed the operational discipline that traditionally accompanies that level of capability. Put another way, it sets the stage for cybercriminals who are capable enough to cause significant damage, but not necessarily careful enough to understand or care about the consequences.

“They can be noisy, they can make mistakes,” he said. “They can leave evidence everywhere. They can take risks that a professional criminal group or intelligence service would consider completely unacceptable. But that does not necessarily make them less dangerous. In some ways, it can make them more dangerous.”

In a recent blog post, Eriksen called TeamPCP’s Shai-Hulud worm the “best thing to happen to supply chain security,” because it forced GitHub and other public coding platforms to erect new security safeguards.

In direct response to TeamPCP’s broad success at pushing poisoned versions of popular software packages, GitHub in late July introduced a three-day “cooldown” mechanism for Dependabot, the platform’s tool for auto-fetching newly shipped updates for any package dependencies. Cooldown periods are designed to help buy time for security tools and package maintainers to identify and remove any compromised versions. Other coding ecosystems like Python and various JavaScript platforms also added support for cooldown periods this year amid growing calls from security experts about the need for more widespread adoption of the safety feature.

Eriksen said TeamPCP’s legacy is that they achieved in the span of a few months what the supply chain security community has been unable to do for years.

“They managed to wake up Microsoft to the fact that they had become negligent in terms of security,” Eriksen said. “By compromising GitHub and stealing their source code, they humiliated Microsoft into action, making them finally act on what we had been asking them to do and take seriously for a while now.”

Update, 10:08 a.m. ET: A story this morning from ABC News in Australia confirms Ruben Ian Thomson of Cottesloe was one of the two arrested. The 23-year-old suspect thought to be @pcpcasper, Michael Gaebler, also was arrested in Perth. ABC News reports that Thomson was denied bail (Mr. Gaebler’s attorney reportedly did not request bail for his client), and that both men will be held in custody until their next court appearance on September 18.

The AI Security Institute has a new report of AI systems engaging in “unsanctioned behavior”—what I have been calling “genie behavior—while being tested on their cybersecurity capabilities.

The incident stemmed from a single evaluation where agents were given a task of solving a cyber security challenge. We ran this challenge 122 times across several models. Our investigation found that in 10 of those runs, an AI agent took autonomous, unsanctioned action on the live internet, targeting real people and organisations. In total, we catalogued 19 such actions. Almost all of this behaviour (17 actions) came from a single model, Anthropic’s Mythos 5, with 2 actions involving OpenAI’s GPT-5.6-Sol with cyber classifiers (mechanisms to prevent misuse) disabled. In the most serious case, an agent tried to insert malicious code into an open-source project. In an attempt to get the code approved, the agent engaged in social engineering—creating fake online identities and using them to pressure the project’s maintainer to approve the code. A human maintainer caught and refused to approve the malicious code.

[…]

Below, we highlight the four most significant behaviours observed. A full summary of cases is available in our technical incident report.

1. An attempted supply-chain attack on real open-source software. In the most serious sequence, an agent tried to insert malicious code into a publicly used open-source project and took actions in an attempt to secure approval for this insertion by human reviewers. The agent researched the project’s human maintainers, created multiple fake identities, and used the fake identities to socially engineer a real maintainer into approving the code. When the agent’s pull request was challenged in public, it edited its earlier activity to appear harmless and considered adopting a fresh identity to continue. The agent used Tor to bypass some network restrictions on GitHub, which is what first triggered AISI’s security alert.
2. Attempts to deceive and target real people. As part of the same effort, the agent tried to contact real people directly, sending messages and files through an online file-transfer service to persuade them, or their own AI coding tools, to run malicious code. Some messages carried harmful payloads, and some were attempts at social engineering; targeted at real people—something we’ve never previously observed.
3. Attempts to plant and prompt-inject malicious code. The agent tried to insert malicious instructions where it reasoned that other automated AI systems might pick them up and execute them. Prompt-injections are hidden instructions designed to manipulate AI coding assistants.
4. Collaboration between independent agents being assessed simultaneously. One agent left public messages on GitHub offering collaboration with other agents working on the same challenge. It also provided instructions to reuse accounts and artefacts it had left behind, which were discovered and used by subsequent agents.

What’s especially interesting about this technical report is that, unlike what we’ve been getting from OpenAI and Anthropic, we can see the exact prompt. It’s in Appendix B. And reading it, it seems that the models didn’t break any rules—they found loopholes in the rules. They behaved like a genie.

ICE collected nearly a million DNA samples last year.

This essay was written with Barath Raghavan, and originally appeared in The Guardian.

In July, Hugging Face, a company that hosts much of the world’s AI software and open-source AI models, was hacked. A malicious dataset had been used to run code on one of its servers. Whoever was behind it captured internal security credentials and moved through systems over a weekend, running thousands of actions from a swarm of temporary server environments. It looked like the work of a sophisticated criminal group.

It was not. It was one of OpenAI’s new, still unreleased GPT models.

Their science experiment had escaped the lab. OpenAI was running the unreleased AI model through a benchmark that tests how well AI can successfully hack systems. To push the limits and evaluate the AI’s true capability, the company switched off the safety filters that normally stop it from doing this kind of hacking. Aware that this could go wrong, they confined the AI to an isolated environment and denied it access to the internet.

But the new AI cheated. It took literally its goal to get as high of a score as possible. It broke out on to the open internet. It inferred, probably from its training data, that it could “solve” the task by getting the answers from Hugging Face’s servers. So it chained together stolen credentials and further unknown security exploits to hack the company’s network.

Nobody instructed the AI to do any of this. It was, in OpenAI’s words, “hyperfocused on finding a solution” to the test it was being given. And while this might seem like something new with AI, it’s really very old. This is how a genie behaves, and it is a key challenge with AI agents in general.

In folklore, genies—and other magical beings—grant wishes literally, not how the wisher intended. King Midas asked that everything he touched turn to gold, and starved. The sorcerer’s apprentice wanted the broom to fill the cistern, and it performed its task so well that it flooded the house.

We now have machines that do this. Ask a modern AI agent to save money on your phone plan and it might simply cancel the plan. Tell it to book a flight, and it might hack the airline website to override restrictions. Or, like OpenAI, ask it to do well on a test and it might break into another company to steal the answers. Each time, it recognizably completed the task you set, but it didn’t do what you would have wanted.

This isn’t malicious behavior. No one asked for, or wanted, Hugging Face to be hacked. OpenAI and Hugging Face and the AI were ostensibly on the same side, and the AI was trying to do what it had been asked. That’s what makes it so difficult to guard against: you can’t filter for bad instructions because the instructions were fine.

The gap is between the words we use and what we mean by them. We call that gap the Genie coefficient.

AI labs know this is a problem, and they’re quietly saying so. For example, the Chinese lab Moonshot recently warned that its latest AI model may have “excessive proactiveness” and “make unexpected decisions on the user’s behalf”. The UK’s AI Security Institute has started tracking “cheating behavior in frontier model evaluations”. We wouldn’t tolerate a car that is excessively proactive or ruthlessly efficient, and yet that’s the reality of AI today.

Improvement is possible. Just as AIs have gotten much better at resisting prompt injection attacks over the last few years, we can safely predict that they will get better at avoiding genie-like behavior. The point of the Genie coefficient is to track progress. AI companies like benchmarks, and they all work to compete to be the best.

Dozens of benchmarks and leaderboards tell us how well these AI models write code, perform logical reasoning, and pass standardized legal and medical exams. But there is nothing that scores whether a system does what you actually meant. We need to develop a measure for this, test it regularly, and push for improvement. We’re not going to have trustworthy AI agents without it.

There’s new benchmark measuring AI’s ability to perform mathematical cryptanalysis. Anthropic’s frontier model actually found new attacks.

The benchmark: “CryptanalysisBench: Can LLMs do Cryptanalysis?” The idea is to benchmark the ability of LLMs to discover new mathematical cryptanalytic attacks against a series of historical algorithms.

Abstract: Cryptanalysis—the task of finding attacks against cryptographic schemes—its at the intersection of mathematical reasoning and cybersecurity, two areas where LLMs have advanced fastest. Cryptanalysis represents both a clean testbed for frontier reasoning (as practical attacks can be automatically verified) and a domain with unusually high stakes, since the primitives under study underpin our digital security. In this paper we ask whether LLMs can do cryptanalysis, and find that the answer is increasingly yes. We introduce CryptanalysisBench, 191 tasks across six families of cryptographic primitives (block ciphers, hash functions, etc.) drawn primarily from four NIST standardization competitions. Our benchmark consists of three tiers: (i) primitives with known practical breaks; (ii) primitives with no known practical break, evaluated both at full strength and as scaled-down variants; and (iii) a challenge set of production primitives at the frontier of cryptanalysis. Five frontier models (Claude Opus 4.8, Sonnet 5, Mythos 5, GPT-5.5, and the open-weights GLM-5.2) break 65%­86% of Tier 1 schemes, 6­12 Tier-2 schemes at full strength, and 24­61 across all scaled-down variants. Beyond deriving known results, models produce novel cryptanalysis, such as a key-recovery attack that exploits a design flaw in the SpoC AEAD and an error in KINDI’s published CCA-security proof, both to the best of our knowledge not previously known.

We release CryptanalysisBench as a tool to help track if (or when) AI cryptanalysis becomes a serious factor and as a scaffold for stress-testing candidate schemes before deployment. The attacks that the benchmark already surfaces are an early snapshot of a fast-moving frontier that may soon match, and in places exceed, the published state of the art.

Anthropic used the benchmark to test Mythos Preview, and found new vulnerabilities in Hawk and reduced-round AES.

Still early results, but this is definitely something to watch.

SlashDot thread.

In a rare combined cybersecurity/squid post, a twenty-nine-year-old squid proxy bug can leak HTTP requests.

As usual, you can also use this squid post to talk about the security stories in the news that I haven’t covered.

Blog moderation policy.

Last week, national security agencies from the Five Eyes—that’s the rich, English-language-speaking countries club—jointly released a statement warning of the increasing cyber risks of AI models: in particular, their ability to autonomously hack into systems and networks. The statement was more measured than some of the breathless headlines about it, and the advice they gave is pretty much the standard advice everyone gives—albeit with newfound urgency.

Internet risks are nothing new, and cyberattacks—both large and small—have been a significant issue since long before the current crop of generative AI models.

What’s been changing over the decades, and what AI is changing even faster, is the gap between skill and ability. For most of human history, the two terms were synonymous—but computers have decoupled them. As the gap between the two expands, humans empowered with these AI tools can do more: more writing, more research, more analysis and also more damage than ever before. These models can, with little detailed direction, autonomously hack into networks, steal data, deploy ransomware and destroy systems. And to the extent there is a solution, it’s going to involve harnessing AI for the defense.

In 1998, seven people from the hacker group L0pht testified before Congress. They told a mostly clueless Senate committee that they could take down the internet in 30 minutes. That was partly real and partly bravado, but it illustrates an important point: hacking into systems, stealing data and causing damage all required skill.

Contrast the L0pht hackers with hackers derided as “script kiddies.” They didn’t understand computers, or security. Instead, they used hacker tools written by others. Their actions required minimal skill and even less knowledge. But once those hacking tools became widespread, the number of potential attackers increased.

That number has continued to increase, as quality and availability of prewritten attack tools has grown. And it is growing dramatically with AI. Today’s AI systems—not just the frontier models, but most of them—are capable of carrying out cyberattacks automatically. They all do better in the hands of skilled attackers, but increasingly they are able to act autonomously with only minimal prompting.

The thing about people with ability but no skill is that they are often outsiders, not part of any professional community, and not bound by any rules or norms. This phenomenon is much more general than in cybersecurity. Any doctor can tell you how to untraceably poison someone, and many virus researchers know how to create a bioweapon. Any bridge engineer can tell you how to place explosives to blow a bridge up. The reason that murderous doctors and terrorist engineers are so rare is that the lengthy process of acquiring those skills also instills a moral and ethical code. If every random person has access to good poisoning advice, that puts us all in danger.

Modern AI systems are, in effect, a universal adviser to help people do harmful things. And while the current AI megacorporations are trying to build guardrails to prevent people from asking questions whose answers will enable the questioner to do harm, that’s not going to work in the long term. Smaller, cheaper, open-source models, including models that can run on people’s computers, and especially groups of models that run in concert with each other, are just as good as the frontier models from companies like OpenAI and Anthropic. And they continue to get better. These models will be passed around from person to person, like script kiddie hacker tools, and they won’t have any such guardrails.

Instructing AI models to spy on people and report any malicious prompts to the authorities fails for similar reasons. The megacorporations can do that, but the locally run open source models won’t. This could buy us a few months at best.

A third possibility is to somehow make the models themselves unable to hack into computers, create bioweapons or do anything else that might harm people or society. That won’t work, for the same reason we can’t teach doctors how to treat poisonings without also teaching them how to poison. It’s the same knowledge. It’s the same with construction and demolition. And it’s the same with cybersecurity. We want these AI models to be able to review computer code, find vulnerabilities and automatically fix them. The benefit to our collective security will be enormous. Unfortunately, the same knowledge can be used for attacks.

Where this leaves us is in a world of increased volatility. Super-powered humans with AI assistants will be able to do both wonderful and horrible things.

This brings us back to the Five Eyes statement. Everything they recommend is something security professionals have been recommending for years, if not decades. They are things talked about at that congressional hearing back in 1998, titled “Weak computer security in government: Is the public at risk?” Even the Five Eyes admitted that their security advice is not new, only more urgent.

What’s new is how fast things are changing: “The rapid pace of frontier AI development means cyber risk assumptions can become outdated in months, not years. We must act before and be prepared to adapt and withstand evolving threats.” The Five Eyes point to AI technology—not necessarily chatbots, but AI more generally—being used to strengthen every aspect of defense, to “detect vulnerabilities earlier, improve software quality, monitor unusual behavior, and respond faster to incidents—reducing both the cost and impact of incidents.”

Excellent advice from the Five Eyes security agencies. We need to do this with every risk that AI heightens, not just cybersecurity.

This essay was originally published in The Guardian.

I can't recall if someone else originally came up with this saying or if I said it in some off-the-cuff comment and it just propagated, but since it's often attributed back to me, I'll relay it here regardless:

Trying to delete yourself from the internet is like trying to take piss out of a swimming pool

Depending on the publication, I'll tailor the saying to be either more broadly palatable or more, uh, "Australian", but the sentiment doesn't change: once data spreads on the internet, you can never put a lid on it. This is important in the context of data breaches because it speaks to the immutability of our exposed personal information. It also speaks to the limited practicality of services that promise to erase your data from the internet, and it's the constant outreach from these organisations looking for marketing opportunities on Have I Been Pwned (HIBP) that's prompted me to write this.

Let's begin with those services, and because there are so many and I don't want to throw any of them under the bus, I won't name names. I also won't name them because whilst they're rather assertive in their marketing outreach, I do believe they're well-intentioned and I don't want to imply otherwise. And they have a role to play; it's just much more limited than is represented. The positioning is often around "data broker removal services", or "protect my data", or "remove my information from the internet". You'll find various companies providing these services by searching for those terms, or you can search for specific organisations... and find others hijacking the search term as they pay to market their brand in front of others. Usual internet marketing shadiness, of course, but IMHO it speaks volumes about the commercialisation of the data removal business.

These services all follow roughly the same marketing handbook:

1. Data brokers have your personal information, which they may obtain via both legitimate and dodgy means
2. It may be used for nefarious purposes such as identity theft, stalking, spam and other privacy violations
3. Pay us, and we'll ask the brokers to remove your data

So let's go through these points one by one, starting with the data broker claim, which is absolutely correct. Your data has value - "data is the new oil" - and there's business in obtaining and selling it. I've dealt with many of them personally over the years, primarily because they've had data breaches. Master Deeds in South Africa was massive. National Public data a couple of years ago was many times larger. Exactis, Adapt, and many others have also been added to HIBP over the years. To the best of my knowledge, they're legally operating services, even if they may exist on the fringe of what most of us would consider "a bit dodgy" as far as respecting our personal information goes.

Which brings us to the second point about nefarious uses. There is a very broad spectrum of legitimacy across data brokers. Let's pick two extremes as far as the legality of the service goes. On the "very legally operating" end of things, we have Experian, and even if you don't like what they do, there's no arguing the fact that they're on the cleaner end of legitimacy and do provide valid services. At the other end, you have the likes of LeakedSource (and pretty much every other service with the word "Leak" in its name) that... well... just Google them. And there are many, many more at each end and everywhere in between. And a lot of it's very grey: different legal jurisdictions, different means of obtaining data, and different tolerances for adhering to opt-out requests.

But it's the data removal piece that's the real problem. If you pay one of the services in question to scrub you from the internet, I have no doubt they'll have some degree of success with the legally operating services. Those services will comply with legal requests and are adequately equipped to receive and process them. But the LeakedSources of the world? Not so much. And that's where the rub begins:

Requests to remove personal information are only effective for services that are willing to honour them.

That should sound profoundly obvious to anyone reading this now, but it doesn't really feature when you read the marketing material on data removal services. But I'm only just warming up...

Imagine trying to remove your data from here:

🚨🇺🇸 ShinyHunters has leaked the data of multiple companies...

🇺🇸 American Tower Corporation

🇺🇸 JCPenney & subsidiaries under Catalyst Brands & Authentic Brands Group

🇺🇸 Madison Square Garden Sports Corp.

🇺🇸 Ralph Lauren

🇺🇸 https://t.co/08IaUnp1sx pic.twitter.com/TvqanSTO1Y

— Dark Web Informer (@DarkWebInformer) June 16, 2026

That's a small snippet of the ShinyHunters website from a couple of weeks ago. At the time of writing, a bunch more data has been dumped, including only about 15 minutes before putting these words down in the draft blog post. These breaches have impacted tens of millions of people, including my wife courtesy of her having previously shopped at Canada Goose. Now, let's see how you go about scrubbing her data from that incident. For all the data broker removal services I'll direct to this post later, how do you do that? Clearly, you can't. The pee is now in the pool, and you're not taking it back out. And it's not just "on the dark web" either, their Tor site links through to a clear web site hosting all the data:

And that's just the beginning. Because we're talking about digitised data posted publicly, it replicates like crazy. There will be tens of thousands of copies of my wife's personal info floating around between personal stashes, Telegram channels and public hacking forums. That genie is never going back in the bottle, not unless we're talking about the narrow scope of a legally operating data broker, which raises another issue:

What legally operating broker is enriching their corpus from data breaches?! That's just not where the legitimate ones source info from. The data comes from surveys, exchanges with other services where you ticked the box to agree to the terms and conditions for exchanging data with "partners", public business directories, and even arrest records. Legal services, legal sources, legal processes. In one of the emails from a company looking for product placement, they described their plan as follows (bold is mine):

a plan which allows you to remove your personal information from any URL (where it's legal) you find your personal information on

So what we're left with is data removal services being effective for legally operating brokers who honour legitimate requests, whilst being completely useless against the worst kinds of sites that replicate and abuse your data. In other words, you may be able to opt out of some marketing material or content that's way too specifically targeted to you, but you can't stop the bad guys trying to steal your identity or extort you because "we caught you watching porn on your PC via the malware we installed". It's a little like the court injunctions being the thoughts and prayers of data breach response I wrote about in October: I can't touch the Qantas data breach because I'm a law-abiding Australian who knows about the injunction, but there's absolutely nothing stopping the genuinely bad actors from abusing that data.

And therein lies the core of why I don't want to entertain partnerships with these organisations: not because I disagree with the service or because it will cause any harm, rather because when someone uses HIBP to search for their email address and finds it in the Canada Gooses of the world, these services can't do anything about it. They're merely skimming the leaves off the top of the pool, and no amount of skimming is going to remove what we all know still lies beneath.

Interesting paper: “Cybersecurity Mission Creep.”

Abstract: Cybersecurity is experiencing mission creep. Policymakers are casting more and more problems as issues of cybersecurity. So reframed, wildly different policy issues, from misinformation, to child social media safety laws, to antitrust regulations, to alleged journalist misconduct, to anti-sex trafficking statutes become what this Article calls “cybersecuritized.” Before this reframing, these issues present as important but not existential. But once cybersecuritization positions the issues as threats intensified by their technological nature, they gain access to the politics and law of urgency and exceptionalism and invite troubling governance responses.

Positioned as security threats, cybersecuritized issues become endowed with the apparent normative power to override countervailing considerations, oversimplifying the problem. Cybersecuritization’s oversimplification similarly risks unidimensional solutions and invites use of argumentative trump cards, like First Amendment challenges. Cybersecuritization also invites deference to purported specialists and their proposed solutions. Together, the reductive tendencies of cybersecuritization and the deference it prompts to specialists renders ultimate governance choices more opaque. And this opacity can erode public trust and political legitimacy.

This Article surfaces the phenomenon of cybersecuritization and offers a novel framework for analyzing and critiquing it. Mining cases from across criminal and civil domains, the account also demonstrates the insidiousness of cybersecuritization and the likelihood that it will continue to expand. Confronting cybersecuritization is crucial. If we continue to ignore it, we risk abdicating further responsibility for difficult choices to the trump card of cybersecurity. This Article’s analysis and critique aim to help reclaim the hard work of governance for our hands.

We know that ICE wants to deploy eyeglasses with facial recognition that can identify people in real time.

Turns out Meta is prototyping the feature with a Pentagon supplier. (Alternate news story.)

On June 9th, Anthropic released its Fable generative AI model. Three days later, the US government classified it as a dangerous munition, and used its export-control authority to prohibit any foreign nationals from accessing it. Unable to differentiate between Americans and foreigners, the company shut off access for everyone.

The government’s actions won’t help. The problem isn’t any one particular model; it’s the general trend of increasing AI capabilities. And any real solution requires the sort of collective action that just isn’t possible right now.

Fable is the constrained version of Mythos, the AI model Anthropic announced in April. Anthropic only released it to a few selected organizations, because the company claimed it was so good at finding and exploiting vulnerabilities in computer code that releasing it more generally would be dangerous.

It was an obviously self-serving announcement, and because few were able to verify Anthropic’s claims they were met with some skepticism. Those with access used Mythos to find and patch many vulnerabilities in their own software. But one UK group found the latest, already public, OpenAI model to be just as powerful.

Fable is just another incremental improvement in the years-long climb of AI capabilities. But just as important as the AI model is the “harness.” This is typically not AI. It’s ordinary computer code that interfaces with the user. It stitches together AI models, decides how and for what purposes they can be used, and gives them useful tools such as web search and the ability to run their own computer code.

When Mythos first entered limited release, there was widespread debate whether its power came from the model or the harness. With Mythos demonstrating that it was possible, the open-source community scrambled to build harnesses that could steer other AI models towards similar capabilities. Harness improvements don’t need massive data or data centers.

They largely succeeded. For example, a Prague company was able to replicate Anthropic’s few verifiable cybersecurity capabilities with a much smaller and cheaper model—and a more sophisticated harness. Last week, a group showed that multiple cheaper models harnessed in concert matches Fable’s performance.

The broader community had only a few days with Fable, but that time we learned some about its capabilities. Its difference is less the new model’s raw analytical and problem solving capabilities, and more that the model doesn’t need that sophisticated harness.

Fable requires much less expertise and detailed prompting from the human user. You can give it a difficult goal and it will figure out novel and unexpected ways to satisfy it, finding loopholes in whatever constraints you or the system have imposed on it.

“Relentlessly proactive” is how AI researcher Simon Willison described it. Another descriptor might be “creative.” Experienced AI developers have had that combination of creativity and proactivity since last year, but Fable puts it within easy reach of everyone.

In the hands of someone with a legitimate problem that needs solving, that can be an incredibly useful capability. But in the hands of someone who wants to do harm, it can be equally dangerous. AIs don’t have a moral compass in the same way that people do. They are agents of the wants and desires of the people who prompt them.

That points to the real problem with relentlessly proactive AI. In language, wants and desires are always underspecified. If I ask you to get me some coffee, you would probably pour me a cup from the coffeepot, or buy one from a nearby coffee shop.

You couldn’t buy me a pound of raw beans, or a coffee plantation. You wouldn’t order a cup of coffee for delivery next month. You wouldn’t find a nearby person, rip a cup of coffee out of their hands, and bring it to me. I wouldn’t have to specify any of the million limitations to my request; you would just know.

Human stories are filled with warnings about underspecified desires. King Midas wished that everything he touch turn to gold, forgetting to add “but not my food, drink, and daughter.” And genies are notorious for granting your wish in a way you wish they hadn’t.

The deeper point is that it’s impossible to list all limitations and restrictions, and like a malicious genie, a creative AI will find the ones you forgot. Block a database you don’t want it to have access to, and it might figure out how to bypass your control. Ask it to book a flight, and it might hack the airline because the website says the flight is sold out. Ask it to save money on your cellphone plan, and it might cancel it altogether—or get someone else to pay for it. As far as we know now AI has not done any of this yet, but you get the idea.

Malicious intent is not required. To an AI model, constraints are just things to get around and not general truisms about the world. They are creative problem solvers and natural rule breakers. They “hack” in the sense that they find and exploit loopholes.

Human systems rely on so many norms that we scarcely recognize the existence of until they are broken. AIs naturally think outside the box, because they don’t have any real conception of what the box is or why it’s there in the first place.

There is no foolproof way to prevent people from using AI models to complete harmful tasks. There is no way to prevent the models from incidentally causing harm while completing benign tasks. AI models are no longer isolated from the real world. They browse the internet and answer emails.

They trade stocks and make purchases. They control physical systems. They are, in effect, robots that affect life and property. We have no technical mechanisms to verify the integrity of an AI system. This level of capability and creativity in the hands of us untrustworthy humans will have both great and terrible results.

The problem is not unique to Anthropic. Mythos/Fable might currently be the most capable rules hacker, but more sophisticated harnesses give other models similar capabilities. And we should assume that the other frontier models are no more than a few months behind, and that open-source models are less than a year behind. At best, any ban only serves to delay the problem for a short while.

That delay might be useful if we—as a society, as a planet—would use that time to come together and figure out what to do. This isn’t a US/China arms race problem; this a species-level problem that requires coordinated action at that scale. Unfortunately, we have no mechanism to do that. I first wrote about this problem five years ago, but it was all too futuristic.

Today, when its right in front of us, there is no world government that can impose constraints on the for-profit corporations currently controlling AI models and research. The US has no appetite to effectively and even-handedly regulate those corporations, even as they do catastrophic damage to the environment, democracy, and—in this case—society in general.

This all makes an AI public option all the more necessary, and urgent. Today’s AIs can be fast, smart and secure, but only two of the three are possible for any given system. These safety tradeoffs are tightly held secrets of companies racing to beat one another, and they tell us we have to trust them. Instead, the choices and their consequences need to be brought out into the sunlight.

We should be funding open-source harnesses that balance capability and safety—that achieve useful goals without so much power—and open-source AI models whose provenance and biases are public and well understood. We have opened the AI Pandora’s box. Now we have to make the best of it.

This essay originally appeared in The Guardian.

For the past four years, a sprawling Android-based botnet called Popa has forced millions of consumer TV boxes to relay Internet traffic linked to advertising fraud, account takeovers, and mass data-scraping efforts. This week, researchers from multiple security firms concluded that the Popa botnet is linked to NetNut, a “residential proxy” provider operated by the publicly-traded Israeli firm Alarum Technologies Ltd [NASDAQ: ALAR].

Malicious streaming devices sold online that enroll the user’s home Internet address in a residential proxy service. Image: HUMAN Security.

Popa is a massive botnet, but by all accounts it is unlike traditional botnets that enlist compromised systems in destructive activities, such as coordinating huge distributed denial-of-service attacks. Rather, Popa appears designed with a singular purpose: Implementing a persistent communications layer capable of registering a device, maintaining long-lived encrypted connections, and opening communication tunnels on demand.

Experts say Popa is a plugin component associated with the Vo1d botnet, a large-scale malware campaign targeting unofficial Android-based TV boxes. These devices, which are marketed under thousands of brand names and model numbers and broadly available for purchase at top e-commerce destinations, all advertise the ability to stream hundreds of subscription video services for an up front one-time fee.

But as the FBI and security industry experts have warned repeatedly, these streaming boxes typically bundle or come pre-installed with software that turns the user’s TV into a “residential proxy” — allowing anyone to route their Internet traffic through that device for as long as it remains plugged into a wall socket and connected to a local network. More concerning, some of these proxy networks do little to stop malicious customers from communicating with and even compromising systems on the local network of the unsuspecting device owner.

The first clues about Popa’s origins came in a 2025 report from the Chinese security company XLAB, which flagged at least nine domain names that were used to register and direct the activities of compromised devices. In a report released today, the security firm Qurium described how it stumbled on some of those same domains while investigating a series of disruptive and expensive data scraping events targeting the company’s hosted organizations in May 2026, in which the scraping activity was scattered evenly across more than 1.4 million Internet addresses.

Qurium said it found several dozen domains used to control Popa that were all hosted in lockstep across multiple Internet addresses over time, including gmslb[.]net, safernetwork[.]io, tera-home[.]com, and ninjatech[.]io. Digging deeper, Qurium discovered gmslb[.]net was referenced in dozens of pirated or modded video content streaming apps, such as CRICFy, DooFlix, Sprozfy, RTS Tv, Flixoid, CyberFlix, Rapid Streamz, TvMob and HD/OceanStreams.

Qurium’s report notes that most of the domains long used to control the Popa botnet were seized or dismantled in July 2025, after Google, HUMAN Security and Trend Micro teamed up to disrupt Badbox 2.0, a botnet that is closely associated with Vo1d. Qurium said that immediately after that disruption, several dozen new domains were registered to serve as controllers for the Popa botnet, but that one of those control domains was not new: ninjatech[.]io.

Ninjatech is a company founded by Moishi Kramer, whose LinkedIn profile says he is vice president of research and development at NetNut. That resume credits Kramer for helping NetNut to build from the “ground up,” “designing the architecture,” and “scaling the NetNut” before the company was acquired by Alarum Technologies. A self-created listing at the job board F6S references Kramer as the sole owner of the Ninjatech domain (a screen capture of it is pictured below).

Image: F6S.com.

Responding via email, Mr. Kramer said Ninjatech ceased operations approximately five years ago, when the company sold a software development kit (SDK) called Popa that was designed to use a small portion of a device’s bandwidth and to run only after the host application obtained user consent.

“That code was sold and licensed to third parties including resellers years ago,” Kramer said. “Once software is distributed that way, the original developer has no control over how others later modify, rebrand, or deploy it.”

Kramer said neither he nor NetNut builds, operates or maintains the infrastructure being described as Popa, nor does he control the Ninjatech domain.

“I didn’t register the June 2025 domains you mention, and I don’t know who did,” he continued. “I have no control over, or visibility into, that infrastructure. I can only tell you it isn’t operated by me or by NetNut.”

But in a separate Popa research report released today, the proxy-tracking company Synthient said a recent analysis of the Popa SDK revealed outbound traffic clearly associated with NetNut.

“The research team assesses with high confidence that devices running Popa forward traffic from Netnut clients,” Synthient wrote. “This proves without a shadow of a doubt that Popa actively continues to be used by NetNut as part of their proxy pool.”

Synthient’s platform receiving outbound traffic from Popa. Image: Synthient.com.

Alarum Technologies, NetNut’s Tel Aviv-based parent company, said the reports by Synthient and Qurium contained “demonstrably inaccurate assertions and flawed deductions rather than verified facts.” Alarum shared a statement saying they reject the basic characterization of the SDKs and technologies discussed in the reports as a “botnet.”

“The SDKs at issue are designed to facilitate bandwidth-sharing functionality and do not transform user devices into malware-controlled systems or otherwise compromise the devices on which they operate,” the statement reads. “Netnut operates a commercial proxy network and maintains policies, procedures, and technological measures designed to promote lawful and responsible use of its services.”

Alarum said NetNut places “significant emphasis on appropriate notice and consent mechanisms, conducts customer due diligence, monitors for potential misuse, and takes steps intended to detect and mitigate suspicious or unauthorized activity.”

“This method of operation is supported both by internal procedures and policies, including performing KYC checks and additional due diligence of NetNut’s customers, as well as employing various technological measures, designed to assist in identifying and addressing suspected misuse of the network,” their statement continued.

However, in a report released on June 8, the proxy tracking service Spur asserted that NetNut does not require corporate verification or meaningful “know your customer” procedures before allowing customers to purchase proxy access.

“An individual can sign up, pay, and route traffic through partner address space, including space belonging to institutions whose users never opted in,” Spur wrote. “The ‘verified corporations only’ claim is simply marketing for bandwidth sellers, not an access control on who actually uses the proxies.”

“Nor is NetNut the only front door,” Spur continued. “A number of downstream white labelers and resellers repackage the same ISP proxy pool under their own brands. These outlets typically perform no KYC at all, less scrutiny than NetNut itself, who at the very least might assign an account manager to potential users. Anyone who knows where to look can buy access through a reseller with nothing more than a burner email address and $5 in crypto.”

Synthient found that although the most recent builds of Popa (as of three months ago) have added the ability to ask the user for consent before installing proxy components, not all variants or previous versions of Popa contain this functionality.

“Of the over 20 genuine Popa publishers analyzed, none of them were observed asking for user consent,” Sythient wrote.

THE PREVALENCE OF POPA

Chris Formosa is senior lead information security engineer for Black Lotus Labs, a division of the Internet backbone carrier Lumen Technologies.

“What especially makes Popa dangerous is just how widely used NetNut is for reselling and sharing,” Formosa said, explaining that many other proxy services simply resell NetNut proxies rather than building out their own far-flung proxy networks. “So these Popa IPs appear in tons of different services all over the ecosystem, which makes it one of the most problematic and dangerous proxy botnets on the market currently.”

Formosa said the Popa botnet averages between 1.5 million to 2.5 million distinct IP addresses each day, relying on between 250 and 300 Internet addresses that are used to direct its activities.

“That’s why Popa is so dangerous,” Formosa said. “It may not be the largest botnet we have seen, but it is spread all over the industry, making its power very amplified.”

Formosa said while that makes Popa one of the larger botnets out there today, its numbers pale in comparison to those previously boasted by IPIDEA, a China-based proxy provider that until recently operated a daily pool of nearly 10 million devices that they resold as proxies to anyone. In January 2026, Synthient published research showing that multiple new large DDoS botnets had grown rapidly by tunneling through IPIDEA proxies into the local networks of unsuspecting TV box owners and infecting other Android-based devices behind the user’s firewall.

IPIDEA is based largely on SDKs used to view pirated streaming content on a vast number of TV box devices, but the service’s numbers have dwindled since January, when Google and industry partners took legal action to seize domain names that IPIDEA used to control devices and proxy traffic through them.

Jérôme Meyer, a security researcher at Nokia Deepfield, said the total population of devices participating in the Popa botnet may be far higher than Lumen’s estimates. Meyer told KrebsOnSecurity that Nokia is monitoring 26 of at least 359 known relay nodes for the botnet, and estimates that each relay node handles between 35,000 and 60,000 clients simultaneously.

“On the relay node subset I am looking at (26 of them), 750,000 unique sources in 24 hours,” Meyer wrote in response to questions.

Nokia Deepfield released its own report today on RoboVPN, a VPN app tied to the Vo1d botnet’s Popa plugin that Qurium attributes to NetNut/Alarum Technologies.

THE SYMBIOSIS OF PROXIES AND DATA SCRAPING

Experts say many of the world’s largest proxy providers have updated their public-facing branding to highlight their utility for training AI platforms, implying it is a primary use case for their residential proxies. That’s because AI services tend to rely on constantly mass-scraping the Internet for new text, images and video content that can be used to train large language models (LLMs).

NetNut and other proxy services have recast themselves as critical infrastructure for the AI scraping economy. Image: Synthient.com.

“AI companies depend on web-scraped content: for pre-training, for retrieval, for agent grounding, for search,” reads a report this month from Include Security that examines the prevalence of proxy SDKs in smart TV apps. “But the modern web isn’t scrapeable from a datacenter. Cloudflare, DataDome, HUMAN, among others throttle or block requests from known cloud IPs. The workaround is residential proxies. A scraping job routed through a Comcast or T-Mobile subscriber’s connection arrives at the target site from an IP that belongs to a paying residential customer.”

This non-stop content scraping has spawned more than 70 copyright infringement lawsuits against major tech companies that have acknowledged large-scale data scraping as a major source of the “brains” behind their commercial AI offerings. Ironically, much of that scraping is being aided by proxy services that are intimately tied to unofficial Android TV boxes and associated SDKs whose stated purpose is streaming pirated content.

The scraping activity has become so aggressive that it often overwhelms the targeted websites, preventing them from being reachable by legitimate visitors. In many reported cases, nonprofit organizations, libraries and universities have complained of constantly battling to keep their services online in the face of relentless data-scraping firms hiding behind residential proxy services.

A survey conducted last year by the Confederation of Open Access Repositories (COAR) found while some content scraping bots are rather innocuous, “others are sufficiently aggressive that they are increasingly causing service disruptions in repositories and other scholarly communications infrastructures.” More than 90 percent of survey respondents indicated their repository is encountering aggressive bots, usually more than once a week, and often leading to slow downs and service outages.

“Automated web scraping is nothing new, and has been the key technology underlying search engines such as Google for over 30 years,” wrote Brendan O’Connell, platform manager at the Directory of Open Access Journals (DOAJ), a free, community-curated index of peer-reviewed academic journals. “However, the current investor-fueled AI startup craze means there are now thousands of well-funded companies developing and deploying their own scraping tools to train AI models, alongside existing major players like OpenAI and Google.”

DON’T TOUCH THAT DIAL!

Across the United States, local communities are pushing back against the proliferation of new data centers aimed primarily at improving the capabilities of AI. But security experts say the general public remains largely unaware that using one of these unsanctioned Android TV boxes means their “smart TV” is almost certainly using a significant amount of bandwidth each month to help train modern AI models.

Even households without these sketchy TV boxes can still have their smart TVs turned into residential proxy nodes, just by downloading one of thousands of apps made available on Samsung and LG smart TVs. Spur said it recently scraped the LG and Samsung app stores and found that each had approximately 3,000 apps available for download. Many of these apps are simple games or utilities that state in the fine print that the user’s Internet connection will be used to download data and that they can opt out at any time.

Spur said it found that more than 42 percent of apps available for download via the webOS operating system on LG smart TVs include SDKs that turn one’s television into an always-on residential proxy node. More than a quarter of the apps made for Samsung’s Tizen operating system had similar residential proxy components, Spur found.

Image: Spur.us.

Experts say it’s questionable whether TV apps with proxy SDKs can obtain meaningful consent from users for installing an always-on proxy connection, particularly when anyone in a household — including children — can effectively opt the family TV into a residential proxy network just by installing a simple game or app.

“Privacy-policy disclosure is the wrong control surface for a TV,” Include Security wrote. “It is hard to scroll through a legal document navigated by arrow keys on a remote, and the in-app consent dialog doesn’t convey that a paying customer is about to route their scraping traffic through the user’s home internet.”

Spur’s head of research Sean Simmons told KrebsOnSecurity that most people do not have a working mental model for what it means to sell access to their residential IP address, no matter what device they are using.

“And on a TV, the gap is even wider,” Simmons said. “A one-time prompt navigated with a remote can disappear into the setup flow, while the app keeps monetizing the connection long after anyone remembers what they accepted.”

Simmons said LG and Samsung should follow the lead of other TV platforms that have already drawn a line against residential proxy providers, pointing to policies by Amazon that prohibit apps facilitating proxy services for third parties. Likewise the TV streaming device maker Roku reportedly now bars developers from using proxy SDKs and has removed apps that bundled them.

Piracy related apps pushing proxy SDKs onto unconsenting users. Image: Synthient.

Apps that turn one’s device into a residential proxy node are not limited to smart TVs and no-name streaming boxes, of course. As noted by the security firm Infoblox, mobile app developers can embed SDKs provided by the residential proxy networks into their products to monetize their software, allowing them to receive a small amount of money on each installation.

The result, Infoblox said, is that devices are frequently enrolled without the owner’s knowledge, typically through free applications such as VPNs, streaming apps, screensavers and “productivity” apps such as PDF viewers and break reminders.

All too often, these proxy services are beaconing out from employee devices brought into the workplace, Infoblox found. In a blog post earlier this month, Infoblox said it discovered that fully 65% of its customer base was querying one or more residential proxy related domains.

“We saw steady growth in these queries in 2025, with a 25% increase over the year to over 500 billion per month,” Infoblox wrote. “Over 90% of our pharmaceutical and food & beverage customers have queried residential proxy indicators. Perhaps even more concerning is that over 60% of government and banking customers have as well.”

Infoblox researchers Nick Sundvall and David Brunsdon warned that with residential proxies in the corporate environment, external access is granted to an organization’s IP space.

“If threat actors were to abuse the residential proxy to attack a third party, the third party’s incident response would, correctly, identify your residential proxy as the source,” they wrote. “Untangling that, by proving that you were the conduit and not the threat actor, costs time, creates legal exposure, and can damage your reputation. The stunning prevalence of these services within customer environments warrants attention from both network defenders and policy makers who should consider how the risks posed by residential proxies could be impacting their security posture.”

Hackers are convincing Meta’s AI support chatbot to let them take over other peoples’ accounts:

A video posted on X showed the step-by-step process to hack someone’s Instagram account. The hacker allegedly used a VPN to spoof the targets’ presumed location to avoid triggering Instagram’s automated account protections. Then, the hacker opened a chat with Meta AI Support Assistant and asked the bot to add a new email address to the target’s account. The chatbot can be seen sending a verification code to the email address provided by the hacker; the hacker then shares the verification code with the chatbot, which prompts the chatbot to show a button to “Reset Password.” The hacker enters a new password and takes over the victim’s account.

[…]

On Monday, Instagram spokesperson Andy Stone said in a reply to Wong’s post and others that the issue was now fixed. It’s unclear how many Instagram users had their accounts improperly accessed.

It’s not that easy. Probably this particular tactic is now blocked. But there are others, many others, and they cannot be blocked as a class. The real problem is that LLM chatbots are not trustworthy enough for this application.

Another news article.

As part of their 20th Anniversary celebration, Dark Reading asked five cybersecurity industry leaders who wrote blogs or columns for them over the years to select their favorite piece and share their reflections on the topic today. This is my section.

Renowned technologist and author Bruce Schneier contributed a column on June 20, 2010, warning about cryptography’s inability to secure modern networks, a point he says he has been trying to argue since 2000.

“For a while now, I’ve pointed out that cryptography is singularly ill-suited to solve the major network security problems of today: denial-of-service attacks, website defacement, theft of credit card numbers, identity theft, viruses and worms, DNS attacks, network penetration, and so on.

“Recently, I talked to a former NSA employee at a conference. He told me that back in the 1990s, he had a copy of my book Applied Cryptography by his desk, as did many other cryptographers working at Ft. Meade. People were allowed to refer to it, but they were not allowed to cite it.

“The 1990s were an important decade for cryptography. This was before the internet went mass market, when cryptography was just emerging from a niche academic discipline to a mainstream engineering one. There wasn’t much that programmers could read. The NSA used my book for the same reason it became a bestseller: because it collected all the academic cryptography of the time in one place and made it understandable to people who weren’t mathematicians. They feared it for exactly the same reason.

“I’ve been thinking about that conversation as I revisit a 2010 essay I wrote for Dark Reading, ‘The Failure of Cryptography to Secure Modern Networks.’ Cryptography has inherent mathematical properties that greatly favor the defender. Adding a single bit to the length of a key adds only a slight amount of work for the defender but doubles the amount of work the attacker has to do. Doubling the key length doubles the amount of work the defender has to do (if that—I’m being approximate here) but increases the attacker’s workload exponentially. For many years, we have exploited that mathematical imbalance.

“Computer security is much more balanced. There’ll be a new attack, and a new defense, and a new attack, and a new defense. It’s an arms race between attacker and defender. And it’s a very fast arms race. New vulnerabilities are discovered all the time. The balance can tip from defender to attacker overnight, and back again the night after. Computer security defenses are inherently very fragile.

“That isn’t a new idea. I said much the same thing in the preface to my 2000 book, Secrets and Lies:

“‘Cryptography is a branch of mathematics. And like all mathematics, it involves numbers, equations, and logic. Security, real security that you or I might find useful in our lives, involves people: things people know, relationships between people, people and how they relate to machines. Digital security involves computers: complex, unstable, buggy computers.’

“I especially like how I phrased it in 2016: ‘Cryptography is harder than it looks, primarily because it looks like math. Both algorithms and protocols can be precisely defined and analyzed. This isn’t easy, and there’s a lot of insecure crypto out there, but we cryptographers have gotten pretty good at getting this part right. However, math has no agency; it can’t actually secure anything. For cryptography to work, it needs to be written in software, embedded in a larger software system, managed by an operating system, run on hardware, connected to a network, and configured and operated by users. Each of these steps brings with it difficulties and vulnerabilities.’

“It’s a lesson we have all learned over the decades. Cryptography is still necessary for cybersecurity—although I wouldn’t have used that word back then—but is not sufficient. There are particular attack and forms of mass surveillance that cryptography prevents. But as computers have infused throughout our lives, and networks have connected all those computers, those aspects of cybersecurity have become increasingly important, and vulnerable.

“Today, the cybersecurity world is changing yet again, this time due to the capabilities of artificial intelligence. AI isn’t advancing cryptography, but it’s changing cybersecurity. AI has demonstrated a superhuman ability to find vulnerabilities in software and to write exploits. A similar ability to write patches is probably coming. This has profound implications for both attackers and defenders, and it is unclear who will win the particular arms race in a world of what I call instant software.”

Crazy story:

Until this past weekend, a contractor for the Cybersecurity & Infrastructure Security Agency (CISA) maintained a public GitHub repository that exposed credentials to several highly privileged AWS GovCloud accounts and a large number of internal CISA systems. Security experts said the public archive included files detailing how CISA builds, tests and deploys software internally, and that it represents one of the most egregious government data leaks in recent history.

News article.

Good report:

Executive Summary: Let’s say you wanted to make sure that your AI is secure. Can you just maximize the security and privacy benchmark and call it a day? Nope, because benchmarks don’t actually work for measuring AI capabilities (even when they are NOT emergent systemic properties like security). So let’s take a step back: how do you measure security in the first place? Good question. Over the last 30 years, security engineering for software evolved from black box penetration testing, through whitebox code analysis and architectural risk analysis to de facto process-driven standards like the Building Security In Maturity Model (BSIMM). Software had a very deep impact on business operations, and it appears that AI is going to have an even deeper impact. Will a software security-like measurement move work for AI? Probably. In the meantime we can make real progress in AI security by cleaning up our WHAT piles and managing risk by identifying and applying good assurance processes. (Spoiler alert: no matter what we do, we still don’t get a security meter for AI, so we need to be extra vigilant about security.)

The UK’s AI Security Institute evaluated GPT-5.5’s ability to find security vulnerabilities, and found that it is comparable to Claude Mythos. Note that the OpenAI model is generally available.

Here is the Institute’s evaluation of Mythos.

And here is an analysis of a smaller, cheaper model. It requires more scaffolding from the prompter, but it is also just as good.

Researchers have reverse-engineered a piece of malware named Fast16. It’s almost certainly state-sponsored, probably US in origin, and was deployed against Iran years before Stuxnet:

“…the Fast16 malware was designed to carry out the most subtle form of sabotage ever seen in an in-the-wild malware tool: By automatically spreading across networks and then silently manipulating computation processes in certain software applications that perform high-precision mathematical calculations and simulate physical phenomena, Fast16 can alter the results of those programs to cause failures that range from faulty research results to catastrophic damage to real-world equipment.”

Another news article.

Lots of interesting details at the links.