Anthropic's Claude Opus 4.6 Discovers 22 Firefox Vulnerabilities in Two-Week Security Audit

Details

• Claude Opus 4.6 identified 22 security vulnerabilities in Mozilla Firefox during a two-week testing period, with 14 classified as high-severity, representing nearly one-fifth of all critical bugs Mozilla patched throughout 2025.
• Of 112 total reports submitted to Mozilla, 22 received CVE designations for security-critical issues; the remaining 90 addressed non-security problems including crashes and logical errors.
• The vulnerabilities affected Firefox's memory storage systems, access boundary conditions, security measures, and JavaScript engine components; Mozilla deployed patches in Firefox version 148 released February 24.
• Claude scanned approximately 6,000 C++ files and discovered a Use After Free vulnerability within twenty minutes of initial exploration, flagging fifty additional crashing inputs before human validation.
• In a separate evaluation, Anthropic tested Claude's ability to exploit discovered vulnerabilities, successfully developing functional exploits in two cases using roughly $4,000 in API credits across several hundred attempts, though only in test environments with security features disabled.

Impact

The discovery underscores a critical inflection point in AI-assisted cybersecurity: frontier models now match or exceed human researchers in vulnerability detection speed and thoroughness, even in extensively audited projects like Firefox. Claude's identification of 14 high-severity bugs in a single month—equaling roughly one-fifth of Firefox's entire 2025 security patch volume—demonstrates that AI vulnerability discovery operates at a scale human teams cannot sustain. This creates immediate practical pressure on open-source maintainers with limited resources, many of whom lack Mozilla's scale and dedicated security personnel. However, Anthropic's own testing reveals a temporary asymmetry: while Claude excels at finding vulnerabilities, its exploitation capabilities remain crude and environment-dependent. The company explicitly warns this gap will narrow, likely within months based on observed progress trajectories. The competitive and security implications are substantial—organizations that fail to harden codebases before AI-driven exploitation matures face exponentially higher breach risk. The window for defenders to act proactively appears finite, making immediate investment in code hardening, memory safety initiatives, and security automation increasingly urgent rather than optional.