TLDR:
- Claude Opus 4.6 found 22 Firefox bugs in 2 weeks, 14 flagged high-severity by Mozilla researchers.
- The 14 high-severity finds equal nearly a fifth of all such Firefox bugs Mozilla fixed in 2025.
- Claude succeeded in building working exploits in only 2 of several hundred automated attempts.
- Anthropic spent roughly $4,000 in API credits testing Claude’s exploit development capabilities.
Anthropic’s Claude Opus 4.6 identified 22 security vulnerabilities inside Firefox in just two weeks. Fourteen of those bugs were classified as high-severity by Mozilla. That figure represents nearly a fifth of all high-severity Firefox flaws remediated throughout 2025.
The findings emerged from a structured research partnership between Anthropic and Mozilla.
Claude AI Uncovers High-Severity Firefox Bugs at Record Speed
The collaboration began as an internal model evaluation.
Anthropic wanted a harder benchmark after Claude Opus 4.5 nearly solved CyberGym, a known security reproduction test. Engineers built a dataset of prior Firefox CVEs and tested whether the model could reproduce them.
Claude Opus 4.6 replicated a high percentage of those historical vulnerabilities. That raised a concern: some CVEs may already have existed in Claude’s training data.
Anthropic then redirected the effort toward finding entirely new bugs in the current Firefox release.
Within twenty minutes of beginning exploration, Claude flagged a Use After Free vulnerability inside Firefox’s JavaScript engine. Three separate Anthropic researchers validated the bug independently.
A bug report, alongside a Claude-authored patch, was filed in Mozilla’s Bugzilla tracker.
By the time that first report was submitted, Claude had already produced fifty additional crashing inputs. Anthropic ultimately scanned nearly 6,000 C++ files and submitted 112 unique reports to Mozilla. Most fixes shipped to users in Firefox 148.0.
Firefox 148 Ships Fixes as AI Exploit Research Raises New Alarms
Mozilla triaged the bulk submissions and encouraged Anthropic to send all findings without manual validation. That approach accelerated the pipeline significantly. Mozilla researchers have since begun testing Claude internally for their own security workflows.
Anthropic also tested whether Claude could move beyond discovery into active exploitation.
Researchers gave Claude access to the reported vulnerabilities and asked it to build working exploits. The goal was to demonstrate a real attack by reading and writing a local file on a target system.
Across several hundred attempts, spending roughly $4,000 in API credits, Claude succeeded in only two cases.
According to Anthropic’s published findings, the model is substantially better at finding bugs than exploiting them. The cost gap between discovery and exploitation runs at least an order of magnitude.
The exploits that did work required a test environment stripped of standard browser security features. Firefox’s sandbox protections were not present.
Anthropic noted that sandbox-escaping vulnerabilities do exist and that Claude’s output represents one component of a broader exploit chain.
Anthropic urged software developers to accelerate secure coding practices. The company also outlined a “task verifier” method, where AI agents check their own fixes against both vulnerability recurrence and regression tests.
Mozilla’s transparent triage process helped shape that approach throughout the research.
The post AI Model Finds 22 Firefox Vulnerabilities in Two Weeks appeared first on Blockonomi.
Source: https://blockonomi.com/ai-model-finds-22-firefox-vulnerabilities-in-two-weeks/
