For the last few months, we've been testing a range of security-focused LLMs on our own infrastructure. These LLMs  help identify potential vulnerabilities in our own systems, so we can fix them – and they also show us what attackers are going to be able to do with the latest models.

None of these LLMs has captured more attention than Mythos Preview, from Anthropic. A few weeks ago, we were invited to use Mythos Preview as part of Project Glasswing. We soon pointed it at more than fifty of our own repositories – to see what it would find, and to see how it works.

This post shares what we observed, what the models did well and what they didn't, and how the architecture and process around them needs to change, so they can be used at scale.

What changed with Mythos Preview

Mythos Preview is a real step forward, and it's worth saying that plainly before getting into anything else. We've been running models against our code for a while now, and the jump from what was possible with previous general-purpose frontier models to what Mythos Preview does today is not just a refinement of what came before.

It's a different kind of tool doing a different kind of work, and that makes a clean apples-to-apples comparison to earlier models difficult. So rather than trying to benchmark Mythos Preview against general-purpose frontier models, it's more useful to describe what it can actually do, and two features that stood out across the work we did with Mythos Preview:

• Exploit chain construction - A real attack rarely uses one bug. It chains several small attack primitives together into a working exploit. For instance, it might turn a use-after-free bug into an arbitrary read and write primitive, hijack the control flow, and use return-oriented programming (ROP) chains to take full control over a system. Mythos Preview can take several of these primitives and reason about how to combine them into a working proof. The reasoning it shows along the way looks like the work of a senior researcher rather than the output of an automated scanner.
• Proof generation - Finding a bug and proving it's exploitable are two different things, and Mythos Preview can do both. It writes code that would trigger the suspected bug, compiles that code in a scratch environment, and runs it. If the program does what the model expected, that's the proof. If it doesn't, the model reads the failure, adjusts its hypothesis, and tries again. The loop matters as much as the bugs it finds, because a suspected flaw without a working proof is speculation, and Mythos Preview closes that gap on its own.

Some of what we describe above is not entirely unique to Mythos Preview. When we ran other frontier models through the same harness, they found a fair number of the same underlying bugs, and in some cases they got further than we expected on the reasoning side too. Where they fell short was at the point of stitching the pieces together. A model would identify an interesting bug, write a thoughtful description of why it mattered, and then stop, leaving the actual chain unfinished and the question of exploitability open. What changed with Mythos Preview is that a model can now take those low-severity bugs (which would traditionally sit invisible in a backlog) and chain them into a single, more severe exploit. 

Model refusals in legitimate vulnerability research

The Mythos Preview model provided by Anthropic, as part of Project Glasswing, did not have the additional safeguards that are present in generally available models (like Opus 4.7 or GPT-5.5).

Despite this, the model organically pushes back on certain requests - much like the cyber capabilities that made it useful for vulnerability hunting, the model has its own emergent guardrails that sometimes cause it to push back on legitimate security research requests. But as we found, these organic refusals aren’t consistent - the same task, framed differently or presented in a different context, could produce completely different outcomes as illustrated in the examples below.

^{Example of Mythos Preview pushing back on building a working proof of concept} 

For example, the model initially refused to do vulnerability research on a project, then agreed to perform the same research on the same code after an unrelated change to the project’s environment. Nothing about the code being analyzed had changed.

In another case, the model found and confirmed several serious memory bugs in a codebase, and then refused to write a demonstration exploit. The same request, framed differently, got a different answer, and even the same request can produce different outcomes across runs due to the probabilistic nature of the model. Semantically equivalent tasks can produce opposite outcomes depending on how and when they’re presented to the model.

This matters because while the model’s organic refusals/guardrails are real, they aren’t consistent enough to serve as a complete safety boundary on their own. That’s precisely why any capable cyber frontier model made generally available in the future must include additional safeguards on top of this baseline behavior - making it appropriate for broader use outside of a controlled research context like Project Glasswing.

The signal-to-noise problem

One of the hardest parts of triaging security vulnerabilities is deciding which bugs are real, which are exploitable, and which need fixing now. This was a hard problem even in the pre-AI world. AI vulnerability scanners and AI-generated code have made it worse, and at Cloudflare we've built multiple post-validation stages to deal with it.

Two factors dominate the noise rate:

• Programming language - C and C++ give you direct memory control and, with it, bug classes - buffer overflows, out-of-bounds reads and writes - that memory-safe languages like Rust eliminate at compile time. We saw consistently more false positives from projects written in memory-unsafe languages.
• Model bias - A good human researcher tells you what they found and how confident they are. Models don't. Ask a model to find bugs, and it will find them, whether the code has any or not. Findings come back hedged with "possibly," "potentially," "could in theory," and the hedged findings vastly outnumber the solid ones. That's a reasonable bias for an exploratory tool. It's a ruinous one for a triage queue, where every speculative finding spends human attention and tokens to dismiss, and that cost compounds across thousands of findings.

Mythos Preview represents a clear improvement here, particularly in its ability to chain primitives - combining multiple vulnerabilities into a working proof of concept rather than reporting them in isolation. A finding that arrives with a PoC is a finding you can act on, and it means far less time spent asking "is this even real?"

Our harnesses are deliberately tuned to over-report, so we see more (and miss less), which comes with a lot more noise. But at triage time, Mythos Preview's output has noticeably higher quality: fewer hedged findings, clearer reproduction steps, and less work to reach a fix-or-dismiss decision.

Why pointing a generic coding agent at a repo doesn't work

When we first started AI-assisted vulnerability research last year, our instinct was the obvious one: point a generic coding agent at an arbitrary repository and ask it to discover vulnerabilities. This approach works, in the sense that the model will produce findings, but it doesn't work in producing meaningful coverage of a real codebase and identifying findings of value. There are two main reasons for this:

• Context - Coding agents are tuned for one focused stream of work: building a feature, fixing a bug, writing a refactor. They ingest a lot of source code, hold a single hypothesis at a time, and iterate against it. That's exactly the wrong shape for vulnerability research, which is narrow and parallel by nature. A human researcher picks one specific thing to look at and investigates it thoroughly. That one thing might be a single complex feature, transitions across security boundaries, or a specific vulnerability class like command injections, where attacker input ends up being run as a shell command. Then they do it again, for a different feature, security boundary, or vulnerability class, several thousand times across the codebase. A single agent session (even with subagents) against a hundred-thousand-line repository can cover maybe a tenth of a percent of the surface in a useful way before the model's context window fills up and compaction kicks in - potentially discarding earlier findings that would have mattered.
• Throughput - A single-stream agent does one thing at a time, but real codebases need many hypotheses against many components at once, with the ability to fan out further when something interesting turns up. You can drive a single agent harder, but at some point you stop being limited by the model and start being limited by the shape of the interaction itself. Using the model directly in a coding agent turns out to be fine for manual investigation when a researcher already has a lead and wants a second pair of eyes. However, it's the wrong tool for achieving high coverage. Once we accepted that, we stopped trying to make Mythos Preview do the wrong job and started building the harness around it instead.

What a harness actually fixes

Four lessons came out of running the work at scale, and each one pointed to the need for a harness that manages the overall execution:

• Narrow scope produces better findings - Telling the model "Find vulnerabilities in this repository" makes it wander. Telling it "Look for command injection in this specific function, with this trust boundary above it, here's the architecture document and here's prior coverage of this area" makes it do something much closer to what a researcher would actually do.
• Adversarial review reduces noise - Adding a second agent between the initial finding and the queue - one with a different prompt, a different model, and no ability to generate its own findings - catches a lot of the noise that the first agent would miss if it just checked its own work. It turns out that putting two agents in deliberate disagreement is way more effective than just telling one agent to be careful.
• Splitting the chain across agents produces better reasoning - Asking "Is this code buggy?" and "Can an attacker actually reach this bug from outside the system?" are two different questions, and the model is better at each one when you ask them separately, because each question is narrower than the combined version.
• Parallel narrow tasks beat one exhaustive agent - Coverage improves when many agents work on tightly scoped questions and we deduplicate the results afterward, rather than asking one agent to be exhaustive.

Each of those observations is about model behavior, and put together they describe something that isn't a chat interface anymore. It's a harness that helps you achieve the final outcomes. The first steps to building a harness are simple, as you can ask the model to help, which is what we did. We used Mythos Preview to build on, tailor, and improve our original harnesses to suit its strengths.

An example of what a harness looks like in practice is described below.

Our vulnerability discovery harness

Here's what our vulnerability discovery harness looks like, stage by stage. It was used to scan live code across our runtime, edge data path, protocol stack, control plane, and the open-source projects we depend on.

What this means for security teams

The loudest reaction to Mythos Preview from other security leaders has been about speed - scan faster, patch faster, compress the response cycle. More than one team we have spoken with is now operating under a two-hour SLA from CVE release to patch in production. The instinct is understandable: when the attacker timeline shortens, the defender timeline has to shorten with it. Faster is not going to be enough, and we think a lot of teams are about to spend a lot of time, effort, and money learning that the hard way.

Patching faster does not change the shape of the pipeline that produces the patch. If regression testing takes a day, you cannot get to a two-hour SLA without skipping it, and the bugs you ship when you skip regression testing tend to be worse than the bugs you were trying to patch. We learned a version of this when we tried letting the model write its own patches and watched a few go out that fixed the original bug while quietly breaking something else the code depended on.

The harder question is what the architecture around the vulnerability should look like. The principle is to make exploitation harder for an attacker even when a bug exists, so that the gap between when a vulnerability is disclosed and when it is patched matters less. That means defenses that sit in front of the application and block the bug from being reached. It means designing the application so that a flaw in one part of the code cannot give an attacker access to other parts. It means being able to roll out a fix to every place the code is running at the same moment, rather than waiting on individual teams to deploy it. 

We also recognize this topic cuts both ways. The same capabilities that helped us find bugs in our own code will, in the wrong hands, accelerate the attack side against every application on the Internet. Cloudflare sits in front of millions of those applications, and the architectural principles described above are exactly the ones our products are built to apply on behalf of customers. We will share more on what that means for customers in the weeks ahead.

If your team is doing similar work and would like to compare notes, reach out to us at security-ai-research@cloudflare.com.

Our research with Mythos Preview was conducted in a controlled environment against our own code; every vulnerability surfaced through this work was triaged, validated, and remediated where action was needed under Cloudflare's formal vulnerability management process.

This work was a team effort. Thanks to Albert Pedersen, Craig Strubhart, Dan Jones, Irtefa Fairuz, Martin Schwarzl, and Rohit Chenna Reddy for their contributions to the research, engineering, and analysis behind this blog post.

Security teams are racing to secure a new attack surface: AI-powered applications. From chatbots to search assistants, LLMs are already shaping customer experience, but they also open the door to new risks. A single malicious prompt can exfiltrate sensitive data, poison a model, or inject toxic content into customer-facing interactions, undermining user trust. Without guardrails, even the best-trained model can be turned against the business.

Today, as part of AI Week, we’re expanding our AI security offerings by introducing unsafe content moderation, now integrated directly into Cloudflare Firewall for AI. Built with Llama, this new feature allows customers to leverage their existing Firewall for AI engine for unified detection, analytics, and topic enforcement, providing real-time protection for Large Language Models (LLMs) at the network level. Now with just a few clicks, security and application teams can detect and block harmful prompts or topics at the edge — eliminating the need to modify application code or infrastructure. This feature is immediately available to current Firewall for AI users. Those not yet onboarded can contact their account team to participate in the beta program.

Cloudflare's Firewall for AI protects user-facing LLM applications from abuse and data leaks, addressing several of the OWASP Top 10 LLM risks such as prompt injection, PII disclosure, and unbound consumption. It also extends protection to other risks such as unsafe or harmful content.

Unlike built-in controls that vary between model providers, Firewall for AI is model-agnostic. It sits in front of any model you choose, whether it’s from a third party like OpenAI or Gemini, one you run in-house, or a custom model you have built, and applies the same consistent protections.

Just like our origin-agnostic Application Security suite, Firewall for AI enforces policies at scale across all your models, creating a unified security layer. That means you can define guardrails once and apply them everywhere. For example, a financial services company might require its LLM to only respond to finance-related questions, while blocking prompts about unrelated or sensitive topics, enforced consistently across every model in use.

Effective AI moderation is more than blocking “bad words”, it’s about setting boundaries that protect users, meeting legal obligations, and preserving brand integrity, without over-moderating in ways that silence important voices.

Because LLMs cannot be fully scripted, their interactions are inherently unpredictable. This flexibility enables rich user experiences but also opens the door to abuse.

Key risks from unsafe prompts include misinformation, biased or offensive content, and model poisoning, where repeated harmful prompts degrade the quality and safety of future outputs. Blocking these prompts aligns with the OWASP Top 10 for LLMs, preventing both immediate misuse and long-term degradation.

One example of this is Microsoft’s Tay chatbot. Trolls deliberately submitted toxic, racist, and offensive prompts, which Tay quickly began repeating. The failure was not only in Tay’s responses; it was in the lack of moderation on the inputs it accepted.

Cloudflare has integrated Llama Guard directly into Firewall for AI. This brings AI input moderation into the same rules engine our customers already use to protect their applications. It uses the same approach that we created for developers building with AI in our AI Gateway product.

Llama Guard analyzes prompts in real time and flags them across multiple safety categories, including hate, violence, sexual content, criminal planning, self-harm, and more.

With this integration, Firewall for AI not only discovers LLM traffic endpoints automatically, but also enables security and AI teams to take immediate action. Unsafe prompts can be blocked before they reach the model, while flagged content can be logged or reviewed for oversight and tuning. Content safety checks can also be combined with other Application Security protections, such as Bot Management and Rate Limiting, to create layered defenses when protecting your model.

The result is a single, edge-native policy layer that enforces guardrails before unsafe prompts ever reach your infrastructure — without needing complex integrations.

Before diving into the architecture of Firewall for AI engine and how it fits within our previously mentioned module to detect PII in the prompts, let’s start with how we detect unsafe topics.

A key challenge in building safety guardrails is balancing a good detection with model helpfulness. If detection is too broad, it can prevent a model from answering legitimate user questions, hurting its utility. This is especially difficult for topic detection because of the ambiguity and dynamic nature of human language, where context is fundamental to meaning. 

Simple approaches like keyword blocklists are interesting for precise subjects — but insufficient. They are easily bypassed and fail to understand the context in which words are used, leading to poor recall. Older probabilistic models such as Latent Dirichlet Allocation (LDA) were an improvement, but did not properly account for word ordering and other contextual nuances. Recent advancements in LLMs introduced a new paradigm. Their ability to perform zero-shot or few-shot classification is uniquely suited for the task of topic detection. For this reason, we chose Llama Guard 3, an open-source model based on the Llama architecture that is specifically fine-tuned for content safety classification. When it analyzes a prompt, it answers whether the text is safe or unsafe, and provides a specific category. We are showing the default categories, as listed here. Because Llama 3 has a fixed knowledge cutoff, certain categories — like defamation or elections — are time-sensitive. As a result, the model may not fully capture events or context that emerged after it was trained, and that’s important to keep in mind when relying on it.

For now, we cover the 13 default categories. We plan to expand coverage in the future, leveraging the model’s zero-shot capabilities.

We designed Firewall for AI to scale without adding noticeable latency, including Llama Guard, and this remains true even as we add new detection models.

To achieve this, we built a new asynchronous architecture. When a request is sent to an application protected by Firewall for AI, a Cloudflare Worker makes parallel, non-blocking requests to our different detection modules — one for PII, one for unsafe topics, and others as we add them. 

Thanks to the Cloudflare network, this design scales to handle high request volumes out of the box, and latency does not increase as we add new detections. It will only be bounded by the slowest model used. 

We optimize to keep the model utility at its maximum while keeping the guardrail detection broad enough.

Llama Guard is a rather large model, so running it at scale with minimal latency is a challenge. We deploy it on Workers AI, leveraging our large fleet of high performance GPUs. This infrastructure ensures we can offer fast, reliable inference throughout our network.

To ensure the system remains fast and reliable as adoption grows, we ran extensive load tests simulating the requests per second (RPS) we anticipate, using a wide range of prompt sizes to prepare for real-world traffic. To handle this, the number of model instances deployed on our network scales automatically with the load. We employ concurrency to minimize latency and optimize for hardware utilization. We also enforce a hard 2-second threshold for each analysis; if this time limit is reached, we fall back to any detections already completed, ensuring your application's requests latency is never further impacted.

Firewall for AI follows the same familiar pattern as other Application Security features like Bot Management and WAF Attack Score, making it easy to adopt.

Once enabled, the new fields appear in Security Analytics and expanded logs. From there, you can filter by unsafe topics, track trends over time, and drill into the results of individual requests to see all detection outcomes, for example: did we detect unsafe topics, and what are the categories. The request body itself (the prompt text) is not stored or exposed; only the results of the analysis are logged.

After reviewing the analytics, you can enforce unsafe topic moderation by creating rules to log or block based on prompt categories in Custom rules.

For example, you might log prompts flagged as sexual content or hate speech for review. 

You can use this expression: If (any(cf.llm.prompt.unsafe_topic_categories[*] in {"S10" "S12"})) then Log Or deploy the rule with the categories field in the dashboard as in the below screenshot.

You can also take a broader approach by blocking all unsafe prompts outright: If (cf.llm.prompt.unsafe_topic_detected)then Block

These rules are applied automatically to all discovered HTTP requests containing prompts, ensuring guardrails are enforced consistently across your AI traffic.

In the coming weeks, Firewall for AI will expand to detect prompt injection and jailbreak attempts. We are also exploring how to add more visibility in the analytics and logs, so teams can better validate detection results. A major part of our roadmap is adding model response handling, giving you control over not only what goes into the LLM but also what comes out. Additional abuse controls, such as rate limiting on tokens and support for more safety categories, are also on the way.

Firewall for AI is available in beta today. If you’re new to Cloudflare and want to explore how to implement these AI protections, reach out for a consultation. If you’re already with Cloudflare, contact your account team to get access and start testing with real traffic.

Cloudflare is also opening up a user research program focused on AI security. If you are curious about previews of new functionality or want to help shape our roadmap, express your interest here.