Anthropic’s Claude AI Uncovers 22 Security Flaws in Firefox

San Francisco: Anthropic’s AI model, Claude Opus 4.6, found 22 security flaws in the Firefox browser during a two‑week project with Mozilla. This showcases how AI tools are becoming increasingly useful in helping experts identify security problems more efficiently. 

According to Mozilla, the discovered issues include 14 high‑severity, 7 moderate, and 1 low‑severity vulnerability. Most of these flaws have already been patched in Firefox version 148, while the remaining ones are scheduled for upcoming releases. The vulnerabilities were uncovered between January and February 2026, when Claude scanned nearly 6,000 C++ files across Firefox’s codebase. 

Claude also made a big discovery early on. Within just 20 minutes, it found a serious bug in Firefox’s JavaScript engine called a use‑after‑free issue. Human experts then tested the bug in a safe, virtual setup to make sure the AI was correct and that it wasn’t making a mistake. Overall, Anthropic sent 112 bug reports to Mozilla. This helped Mozilla review and fix problems much faster, speeding up the entire security‑patching process for Firefox. 

Anthropic also checked whether Claude could actually use the bugs it found to create real attacks. They tried this several times and spent about $4,000 on API credits. In the end, Claude managed to turn only two of the vulnerabilities into working exploits. Researchers say this proves that finding bugs is much easier than turning them into real‑world attacks, even for advanced AI systems.