Mozilla · 2026-05-07 · major
Mozilla Used Claude Mythos to Find 271 Firefox Vulnerabilities — Including a 20-Year-Old XSLT Bug and 15-Year-Old <legend> Flaw
Firefox engineers built an agentic fuzzing harness around Claude Opus 4.6 then Mythos Preview, surfacing 180 sec-high vulnerabilities — sandbox escapes, UAFs, IPC races — for Firefox 150's 423-fix April release.
Mozilla engineers point Anthropic's Mythos Preview at Firefox's fuzzing harness and ship 271 vulnerability fixes, dwarfing prior years' totals.
Key specs
| Bugs found | 271 |
|---|---|
| Sec high | 180 |
| Sec moderate | 80 |
| Sec low | 11 |
| April 2026 fixes | 423 |
| April 2025 fixes | 31 |
| Model used | Claude Opus 4.6 → Mythos Preview |
What is it?
A behind-the-scenes report from Mozilla's Brian Grinstead, Christian Holler, and Frederik Braun on a months-long collaboration with Anthropic. Starting with Claude Opus 4.6 and later upgrading to Claude Mythos Preview, the team built an agentic harness on top of Firefox's existing fuzzing infrastructure that drove the model to look for sandbox escapes, use-after-frees, race conditions in inter-process communication, and bugs in legacy code.
How does it work?
Rather than letting the model emit vulnerability reports unchecked, the harness wraps Claude with a fuzzer-style verification loop: candidate findings are dynamically tested in a real Firefox build, eliminating the false positives that crippled previous static-analysis approaches. Engineers still write and review every patch by hand — AI-generated code is reference material, not committed code. Examples published include a pair of unusual sandbox escapes, a 20-year-old XSLT vulnerability, and a 15-year-old bug in the <legend> element parser.
Why does it matter?
Firefox 150's April release shipped 423 security fixes; the same month a year ago shipped 31. Of the 271 bugs attributed directly to the Mythos pipeline, 180 are sec-high — historically a pace human researchers cannot match on a codebase this old. Sandbox escapes alone typically command $20K Mozilla bounties, so this rewrites the economics of browser security research. It is also the most concrete production deployment yet of AI-driven offensive-style security inside a major open-source project.
Who is it for?
Security engineers, browser developers, AI-for-security researchers