Anthropic partnered with Mozilla to audit Firefox's codebase, using Claude Opus 4.6 to discover 22 vulnerabilities over a two-week period. Mozilla assigned 14 of these findings as high-severity vulnerabilities—representing nearly a fifth of all high-severity Firefox flaws remediated in 2025. Most of the identified security issues were immediately patched for hundreds of millions of users in Firefox 148.0.
Automated Bug Hunting at Scale
Anthropic's research team initially targeted Firefox's JavaScript engine because it isolates untrusted external code and presents a critical attack surface. Claude Opus 4.6 identified a Use After Free memory vulnerability within 20 minutes of exploration, allowing researchers to generate a description and proposed patch. Expanding the audit across nearly 6,000 C++ files, Anthropic submitted 112 unique bug reports to Mozilla's Bugzilla issue tracker.
- 22 total vulnerabilities discovered across two weeks of automated testing.
- 14 high-severity bugs validated and remediated by Mozilla engineers.
- 112 unique reports submitted following bulk scanning of C++ source files.
- $4,000 spent in API credits during subsequent exploit-generation evaluations.
Exploitation Limits and Safety Caveats
To test the limits of AI-assisted attacks, Anthropic evaluated whether Claude could write functional exploits for the discovered bugs. Across hundreds of trials costing $4,000 in API credits, the model successfully generated crude exploits in only two cases. Crucially, these exploits only functioned in testing environments where browser sandbox protections were intentionally disabled, proving that defensive mitigations like defense-in-depth remain effective against current AI models.
Verification Tools Accelerate Patches
Anthropic attributes the success of the audit to task verifiers—automated systems that allow AI agents to check their own work in real time. These verifiers test whether a vulnerability is removed while running test suites to catch functional regressions. Mozilla engineers noted that receiving candidate patches, minimal test cases, and detailed proofs-of-concept made verifying AI-generated submissions practical.
Why it matters
This partnership shows that frontier language models currently offer a distinct advantage to defenders, finding complex memory flaws faster than human auditing alone. As model capabilities evolve toward easier exploit generation, establishing fast, verifier-backed patching workflows will be vital to securing critical open-source software before attackers gain the upper hand.