Claude finds 22 security vulnerabilities in Firefox

What this is about

Anthropic published details of a collaboration with Mozilla on October 23, 2026. In February, Claude Opus 4.6 examined about 6,000 C++ files from Firefox and submitted 112 suspicious cases. Mozilla confirmed 22 security vulnerabilities, 14 of which were rated high severity. Most fixes reached users in Firefox 148.

The story is not only the number of findings. In a deliberately weakened test environment, the model also turned two of the discovered bugs into crude exploits. That moves the boundary between automated bug discovery and automated exploitation, although it did not produce an attack against a normal browser.

What Claude actually did

Anthropic first asked the model to reproduce known Firefox bugs in older versions. Claude then searched current source code for new problems. It reported a use after free issue in the JavaScript engine after about 20 minutes. While humans validated that finding, it generated more crashing test cases.

For the exploit evaluation, Claude received a virtual machine, known vulnerabilities and a verifier. Its goal was to read a local file and write the same contents elsewhere. About 350 attempts and roughly $4,000 in API credits produced successful results for two bugs. One was CVE-2026-2796, a type confusion issue at the boundary between WebAssembly and JavaScript.

Why it matters

Browsers constantly process untrusted websites and scripts. A flaw in a JavaScript engine can therefore affect hundreds of millions of people. Mozilla defines a high severity flaw as one that may gather sensitive data or inject data or code into other sites without requiring more than normal browsing actions.

For defenders, the speed is valuable: a model can search large amounts of source code and test cases while specialists confirm relevant findings and validate patches. The same falling cost may eventually help attackers. Anthropic says Claude is currently much better at finding and fixing flaws than at fully exploiting them. That window is useful, but it is not guaranteed to last.

In plain language

Imagine a large apartment building with thousands of doors. Claude does not merely walk past them; it systematically tests each lock and records unusual behavior. The caretaker confirms that 22 locks are genuinely defective. In a training building, Claude can demonstrate entry through two doors. In the real building, however, it still cannot cross the additional security barriers.

A practical example

A browser team receives 100 automatically generated crash reports. A human first reviews ten cases with a reproducible test and a clear cause. Three are real security flaws. For every proposed patch, the team runs two checks: the original bug must no longer occur, and 20,000 existing tests must still pass. The model accelerates discovery, while people retain control over approval, priority and release.

Scope and limits

First, the two exploits ran in a test environment without important defenses such as the browser sandbox. They were not complete attacks against a standard Firefox installation.

Second, performance and cost figures come from Anthropic's own study. Mozilla's collaboration confirms the bugs and severity ratings, but it is not an independent comparison with other models.

Third, Claude produced 112 reports and only a subset became confirmed vulnerabilities. Automated findings create review work and should not be accepted without reproducible tests, root cause analysis and human oversight.

SEO & GEO keywords

Claude Opus 4.6, Firefox, Mozilla, Anthropic, browser security, CVE-2026-2796, WebAssembly, JavaScript, security vulnerabilities, exploit development