Partnering with Mozilla to Strengthen Firefox Security

In collaboration with Mozilla, Claude Opus 4.6 found 22 vulnerabilities in two weeks, of which 14 were classified as high-severity. This is nearly one-fifth of all high-severity vulnerabilities fixed in Firefox in 2025.

Anthropic first tested the model's ability to reproduce past Firefox CVEs, then expanded the scope toward finding new vulnerabilities in the current version of Firefox. In particular, it started from the JavaScript engine and expanded into other areas, and the first vulnerability verified was a Use After Free.

Even before verification was complete, Claude had already found more than 50 additional unique crashing inputs. After technical discussions with Mozilla, the approach shifted from obsessing over individual verification to submitting meaningful reports in bulk, ultimately scanning about 6,000 C++ files and producing 112 unique reports.

Mozilla accepted many of these reports and reflected patches in Firefox 148.0, with the rest to be fixed in subsequent releases. Anthropic explained that to increase confidence when reporting, it focused on the following three things:

• minimal test case
• detailed proof-of-concept
• candidate patch

The same work was also extended to exploits, but even after several hundred attempts and about $4,000 in API credits, only 2 cases actually led to a real exploit. In other words, Claude is currently much stronger at vulnerability detection and patching, while exploitation remains much harder.

However, since some security features were removed in the test environment, the results are closer to the level of a crude exploit. Even so, the fact that AI can take the first steps toward automatically exploiting browser vulnerabilities is presented as a signal that defenders need to further increase their response speed.

Anthropic emphasizes that the task verifier was key in this process. Having a verifier that automatically checks whether a bug has actually been eliminated and whether existing functionality remains unbroken greatly improves the quality of LLM-based patching agents.