Mozilla has quietly turned one of the world’s most controversial AI models into a defensive weapon, using Anthropic’s Claude Mythos to help find and fix 271 security vulnerabilities in Firefox 150 before the browser shipped to hundreds of millions of users. The experiment offers an early, concrete glimpse of how so‑called “frontier” AI systems may reshape cybersecurity, boosting defenders’ speed and reach while also lowering the barrier for attackers.
What Mozilla actually did with Mythos
Mozilla’s work with Anthropic began earlier this year before Mythos ever touched Firefox.
In February, Anthropic’s “Frontier Red Team” approached Firefox engineers after using Claude Opus 4.6 to scan about 6,000 C++ files in the browser’s codebase. That initial pass produced 112 unique bug reports, including 22 confirmed security vulnerabilities that Mozilla fixed in Firefox 148, 14 of them rated high severity.
Encouraged by that result, Mozilla then gained access to Claude Mythos Preview, Anthropic’s unreleased, security‑focused frontier model distributed under the restricted Project Glasswing program. For Firefox 150, Mythos was pointed at much larger swaths of the browser’s C++ and Rust code, with the aim of systematically surfacing memory‑safety issues, logic bugs and edge‑case vulnerabilities.
The result: 271 discrete vulnerabilities identified and fixed in the run‑up to the Firefox 150 release.
Mozilla’s Chief Technology Officer Bobby Holley called the experience “vertigo‑inducing,” adding that Mythos essentially gave his team a way to run a top‑tier security researcher at superhuman speed through decades of accumulated browser code.
Why the numbers don’t match the CVE list
Firefox 150’s official security advisory, MFSA 2026‑30, lists 41 CVEs, including three catch‑all memory‑safety entries that group many bugs under a handful of identifiers. That led some observers to ask: if Mythos found 271 issues, where are they?
Mozilla explains that the 271 figure counts individual code defects, many of which were bundled together into those aggregate CVEs, a common practice when closely related memory‑safety bugs share the same underlying cause. The CVE list reflects how vulnerabilities are tracked and disclosed industry‑wide; the bigger number reflects what the AI actually turned up in the code.
In other words, the AI wasn’t inventing hundreds of exotic new classes of bugs. It was exhaustively surfacing variations on well‑understood patterns, use‑after‑free, buffer overflows, logic errors in edge‑case checks, at a scale and speed human teams could not match.
“So far we’ve found no category or complexity of vulnerability that humans can find that this model can’t,” Mozilla told Engadget. “The point is not that Mythos is discovering magic new attack classes, it’s that we can finally find all the old ones.”
How Mythos works as a “super‑analyst”
Anthropic pitches Mythos as an AI system tuned for code reasoning and cybersecurity, capable of scanning vast codebases, understanding context, and even generating working proof‑of‑concept exploits.
Mozilla’s experience backs up that framing:
- Bug reports from Anthropic’s team included minimal, high‑quality test cases, making it easy for Firefox engineers to reproduce and validate issues.
- Mythos could follow complex control flow through multiple files and layers of abstraction, rather than just spotting obvious anti‑patterns in a single function.
- The AI did not seem “confined” to one language: it handled C++ and Rust, and Anthropic has separately demonstrated Mythos finding a 27‑year‑old vulnerability in OpenBSD, long considered one of the most hardened operating systems.
Holley’s takeaway, as quoted by The Next Web and others, is that Mythos behaves like “an elite security engineer who doesn’t sleep and doesn’t get bored,” but still needs humans to triage, prioritize and actually ship fixes.
Why Mozilla says “the zero‑days are numbered”
In a blog post titled “The zero‑days are numbered”, Mozilla argued that Mythos‑class tools could shift the balance of power toward defenders by compressing years of audit work into weeks.
So far, the Firefox team says:
- Mythos has not found any fundamentally new bug type beyond human capability; everything it reports is something a very skilled human could, in theory, have found.
- The advantage is scale and speed: 271 issues in one evaluation pass, across more than two decades of accumulated browser code, is something no human team would realistically attempt, much less complete before a release.
- That suggests there is a finite reservoir of legacy vulnerabilities in mature software, and tools like Mythos finally give defenders a way to drain it, provided they invest time and people into turning AI findings into patches.
SecurityWeek notes that Mozilla sees Mythos as a proof that “AI‑augmented defense” can harden widely deployed software before comparable tools inevitably reach criminal hands.
The dual‑use dilemma: if Mythos can do it, others will too
The same properties that make Mythos a powerful defensive scanner make it a frightening offensive tool.
The UK AI Security Institute, which has evaluated Mythos, says the model can autonomously carry out multi‑stage network attacks, from reconnaissance through exploit development, across common operating systems and services. Anthropic itself has warned that Mythos is so effective at finding vulnerabilities that releasing it fully open would be irresponsible without giving critical infrastructure operators time to patch.
Mozilla is not part of Anthropic’s high‑profile Project Glasswing early‑access club, but it leveraged an existing relationship with the company to get Mythos Preview on its code. Privacy Guides notes that this isn’t Mozilla’s first AI experiment: in March, Anthropic’s earlier Frontier Red Team submitted about a dozen security discoveries that shipped as fixes in Firefox 148.
The difference with Mythos is the order of magnitude. If one model can surface hundreds of bugs in a browser in a single run, similar tools, whether from Anthropic, competitors or open‑source efforts, could be turned on VPN appliances, routers, banking systems or anything else connected to the internet.
That dual‑use tension, Mythos as both shield and potential sword, is what makes the Firefox case more than a clever engineering story.
What this means for users and for the tech industry
For everyday Firefox users, the immediate impact is simple: update to version 150. It includes patches for:
- 41 documented CVEs, including three large memory‑safety roll‑ups.
- A total of 271 AI‑assisted bug fixes, many of them security‑sensitive.
ZDNet notes that Firefox 150 also brings user‑facing improvements, enhanced split view, tab‑sharing tweaks, but the sheer number of patched vulnerabilities makes it one of the most significant security releases in the browser’s history.
For the broader industry, the episode points to several emerging norms:
- AI‑assisted code review and red‑teaming are likely to become standard for major projects, especially those written in memory‑unsafe languages.
- Vendors will need robust triage pipelines: Mozilla had to mobilize multiple engineering teams to validate Mythos reports and avoid drowning in noise.
- Policymakers and regulators will have to grapple with when and how to open up tools like Mythos, balancing the benefits of broad defensive use against the risks of democratized attack capabilities.
As Firefox’s CTO put it, Mythos didn’t change what kinds of bugs exist, it changed how many of them defenders can realistically hope to find. The question now is whether the rest of the software ecosystem will move quickly enough to run the same race before the AI‑powered attackers arrive at the starting line.
