Anthropic says Claude found 22 Firefox vulnerabilities in two-week Mozilla engagement

Anthropic wrote on X on March 6, 2026 that frontier models are now world-class vulnerability researchers, even if they are not yet equally effective at exploiting the bugs they find. The post linked to an official write-up about Anthropic’s collaboration with Mozilla on Firefox security. According to Anthropic, Claude Opus 4.6 discovered 22 Firefox vulnerabilities over two weeks, 14 of them high severity. Anthropic also said that total represented about one-fifth of all high-severity bugs Mozilla remediated during 2025.

Those numbers make the announcement more than a generic benchmark claim. The partnership used a real, large, widely deployed codebase and measured output against bugs Mozilla considered worth remediating. That is a more operationally meaningful result than synthetic capture-the-flag tasks or abstract cyber scoring. Anthropic’s framing is defensive: the company says current models are stronger at finding weaknesses than chaining them into reliable exploits, and it explicitly urges software teams to use that temporary asymmetry to improve security faster.

• Scope: a structured collaboration with Mozilla focused on Firefox.
• Result: 22 vulnerabilities in two weeks, including 14 high-severity findings.
• Claim: today’s frontier models can materially accelerate defensive security work, but offensive capability may continue to improve.

The policy implication is important. Debate around AI and cybersecurity often centers on future offensive risk, but this case shows that defensive use is already operational. A model that can read large codebases, reason across project context, and surface candidate vulnerabilities can change the economics of security review long before fully autonomous exploitation becomes reliable. For maintainers with limited headcount, faster identification and validation of bugs may be the immediate value.

At the same time, Anthropic’s warning matters. The company is not claiming that the current balance will hold. Its message is that defenders have a lead that should be treated as perishable. That makes the Mozilla collaboration significant both as a research milestone and as a timing signal: organizations that wait to modernize their secure development processes may lose the advantage of a period when AI is disproportionately useful for finding and fixing issues rather than weaponizing them.