Anthropic on Friday stated it found 22 new safety vulnerabilities within the Firefox internet browser as a part of a safety partnership with Mozilla.
Of those, 14 have been categorized as excessive, seven have been categorized as average, and one has been rated low in severity. The problems have been addressed in Firefox 148, launched late final month. The vulnerabilities have been recognized over a two-week interval in January 2026.
The substitute intelligence (AI) firm stated the variety of high-severity bugs recognized by its Claude Opus 4.6 giant language mannequin (LLM) represents “virtually a fifth” of all high-severity vulnerabilities that have been patched in Firefox in 2025.
Anthropic stated the LLM detected a use-after-free bug within the browser’s JavaScript after “simply” 20 minutes of exploration, which was then validated by a human researcher in a virtualized setting to rule out the opportunity of a false constructive.
“By the top of this effort, we had scanned almost 6,000 C++ recordsdata and submitted a complete of 112 distinctive stories, together with the high- and moderate-severity vulnerabilities talked about above,” the corporate stated. “Most points have been fastened in Firefox 148, with the rest to be fastened in upcoming releases.”
The AI upstart stated it additionally fed its Claude mannequin entry to your entire listing of vulnerabilities submitted to Mozilla and tasked the AI instrument with growing a sensible exploit for them.
Regardless of finishing up the take a look at a number of hundred instances and spending about $4,000 in API credit, the corporate stated Claude Opus 4.6 was in a position to flip the safety defect into an exploit solely in two instances.
This conduct, the corporate added, signaled two vital features: the price of figuring out vulnerabilities is cheaper than creating an exploit for them, and the mannequin is best at discovering points than at exploiting them.
“Nonetheless, the truth that Claude might succeed at mechanically growing a crude browser exploit, even when solely in a number of instances, is regarding,” Anthropic emphasised, including the exploits solely labored inside the confines of its testing setting, which has had some security measures like sandboxing deliberately stripped off.
An important part included into the method is a job verifier to find out if the exploit truly works, giving the instrument real-time suggestions because it explores the codebase in query and permitting it to iterate its outcomes till a profitable exploit is devised.
One such exploit Claude wrote was for CVE-2026-2796 (CVSS rating: 9.8), which has been described as a just-in-time (JIT) miscompilation within the JavaScript WebAssembly part.
The disclosure comes weeks after the corporate launched Claude Code Safety in a restricted analysis preview as a option to repair vulnerabilities utilizing an AI agent.
“We will not assure that each one agent-generated patches that cross these assessments are ok to merge instantly,” Anthropic stated. “However job verifiers give us elevated confidence that the produced patch will repair the precise vulnerability whereas preserving program performance—and subsequently obtain what’s thought of to be the minimal requirement for a believable patch.”
Mozilla, in a coordinated announcement, stated the AI-assisted method has found 90 different bugs, most of which have been fastened. These consisted of assertion failures that overlapped with points historically discovered by fuzzing and distinct courses of logic errors that the fuzzers didn’t catch.
“The dimensions of findings displays the ability of mixing rigorous engineering with new evaluation instruments for steady enchancment,” the browser maker stated. “We view this as clear proof that large-scale, AI-assisted evaluation is a strong new addition to safety engineers’ toolbox.”
