# Anthropic and Mozilla disclose Claude-assisted Firefox vulnerability work > Claude Opus 4.6 found 22 Firefox vulnerabilities, while Mozilla's triage showed why reproducible cases and coordinated disclosure remain essential. Canonical URL: https://www.devobs.io/news/news-anthropic-mozilla-firefox-vulnerability-collaboration/ By: Lena Fischer Published: 2026-09-06T11:58:54.626Z Updated: 2026-09-06T11:58:54.626Z Event date: 2026-03-06 Section: Security Anthropic described a two-week security collaboration with Mozilla on March 6 in which Claude Opus 4.6 found 22 Firefox vulnerabilities. The [joint-work account](https://www.anthropic.com/news/mozilla-firefox-security) says Mozilla classified 14 as high severity and shipped fixes for most reported issues in Firefox 148, with remaining fixes planned for later releases. ## Discovery scaled faster than human triage Anthropic first tested Claude against historical Firefox CVEs, then searched current code for novel faults. Work began in the JavaScript engine and expanded across nearly 6,000 C++ files. The company says it submitted 112 unique reports after Mozilla asked for bulk reporting, including uncertain crashing inputs, rather than requiring Anthropic to validate every case first. That workflow depended on active maintainer coordination. At large volume, an unstructured stream of model findings can become denial of service for a project. Anthropic reports that Mozilla valued minimal test cases, detailed proofs of concept, and candidate patches. Those artifacts let maintainers reproduce, prioritize, and fix a report without trusting the model's explanation. ## Exploit tests exposed a capability gap Anthropic also asked Claude to exploit submitted vulnerabilities in a reduced-security test environment. It reports success in two cases after several hundred runs and stresses that the crude exploits did not defeat the normal browser sandbox. The result does not establish an end-to-end Firefox compromise; it shows why finding and exploitation must be measured separately. Teams adopting AI vulnerability discovery should establish coordinated-disclosure contacts before scanning third-party code. Gate every submission on a reproducer, affected revision, impact hypothesis, and duplicate search. For patches, run a verifier that confirms the original fault is gone and the existing test suite still passes. The engineering bottleneck is now a trustworthy find-triage-fix pipeline, not raw crash generation. ## Source references -