Claude Opus 4.6 Discovers 22 Firefox Vulnerabilities in Two Weeks — 14 High-Severity

Anthropic and Mozilla ran a coordinated two-week security research project in February 2026 where Claude Opus 4.6 scanned roughly 6,000 Firefox C++ files, submitted 112 reports, and identified 22 CVEs. Fourteen were classified high-severity, representing nearly one-fifth of all critical Firefox bugs patched in 2025. Source

The standout result: a JavaScript engine use-after-free was found in just 20 minutes. The entire operation cost $4,000 in API credits. The vulnerabilities shipped in Firefox 148.0, patching hundreds of millions of users.

There's an important asymmetry in the data. Exploitation succeeded in only 2 of hundreds of attempts — meaning AI-assisted discovery currently outpaces AI-assisted exploitation by a wide margin. That's the right side of the asymmetry for defenders, but it won't last forever.

This result should be read alongside the CTI-REALM benchmark from Microsoft Security AI, which tested 16 frontier models on security detection rule generation and placed Claude Opus 4.6 High at the top (reward 0.637), ahead of Claude Opus 4.5 (0.624) and GPT-5. The convergence is clear: Claude is becoming the default model for security research workflows.

The cost-to-impact ratio here is the real story. $4,000 to find 22 CVEs including 14 high-severity bugs across a codebase serving hundreds of millions of users. No human security team achieves that economics. The question is no longer whether AI-assisted security research works — it's how to operationalize it at scale without creating new attack surfaces in the process.