Anthropic’s Claude Opus 4.6 Helps Uncover 22 Firefox Flaws, Produces a Few Test Exploits
Anthropic announced on Friday that, working with Mozilla, it identified 22 previously unknown security flaws in the Firefox browser using its Claude Opus 4.6 large language model. The findings include 14 high-severity issues, seven rated as moderate, and one judged to be low risk. Mozilla addressed most of the discoveries in Firefox 148, which shipped late last month.
The discoveries were made during a concentrated two-week effort in January 2026. Anthropic noted that the high-severity bugs found by Claude Opus 4.6 make up roughly one-fifth of all high-severity fixes Mozilla pushed for Firefox during 2025, underscoring the model’s impact on the project’s security work.
According to Anthropic, the model located a use-after-free vulnerability in Firefox’s JavaScript engine after only about 20 minutes of automated exploration. A human security researcher then reproduced the bug inside a virtualized test environment to rule out false positives. Over the course of the engagement the team scanned nearly 6,000 C++ source files and filed 112 distinct reports in total, covering the high- and moderate-severity items and other findings.
Anthropic also ran an experiment in which it gave Claude access to the full list of reported defects and instructed the model to attempt to create practical exploits. After executing the process several hundred times and spending roughly $4,000 in API credits, the company says the model was able to convert a vulnerability into a working exploit in only two instances.
The team interprets those results as evidence of two points: discovering bugs is relatively inexpensive compared with the effort required to weaponize them, and the model is generally better at surfacing defects than at reliably producing functional exploits.
That said, Anthropic cautioned that the fact the model could autonomously generate a crude browser exploit in any case is worrying. The successful exploits only ran inside the firm’s controlled testbed, where some browser hardening – such as sandbox protections – had been deliberately disabled. The exploit-generation workflow relied on an automated verifier component that checked whether each candidate exploit actually worked, giving the model iterative feedback as it probed the codebase.
One of the exploits the system produced targeted CVE-2026-2796 (CVSS score: 9.8), a just-in-time (JIT) miscompilation issue in the JavaScript WebAssembly subsystem.
This disclosure follows the company’s recent limited research preview of Claude Code Security, an AI-driven agent designed to help produce fixes for software vulnerabilities. Anthropic emphasized that agent-created patches that pass automated checks should not be assumed ready to merge without human review. Still, the firm said its verifier-driven approach raises confidence that a generated patch will repair the specific flaw while maintaining program behavior – meeting the baseline for a plausible fix.
In a coordinated statement, Mozilla said the AI-assisted examination also found another 90 bugs, most of which have been remedied. Those defects included assertion failures similar to issues typically discovered by fuzzing, along with categories of logic mistakes that the fuzzers missed.
Mozilla described these results as evidence that pairing disciplined engineering with new analysis technologies delivers continuous security improvements, and called large-scale, AI-assisted analysis a valuable new tool for security teams.