AI Discovers 22 Firefox Vulnerabilities in Two Weeks
How an AI Model Outpaced Humans in Cybersecurity
In March 2026, Anthropic’s Claude Opus 4.6 made headlines by identifying 22 security vulnerabilities in Mozilla Firefox within just two weeks. Fourteen of these were classified as high-severity—a staggering 20% of all high-severity Firefox bugs patched in 2025. This breakthrough marks a turning point in cybersecurity, proving that AI can now detect flaws in battle-tested software faster than human researchers.
Mozilla validated the findings and released fixes in Firefox 148. The speed and precision of this discovery raise urgent questions: How did an AI model achieve what traditional methods like fuzzing and static analysis missed? And what does this mean for the future of software security?
Key Findings from the AI-Driven Discovery
Anthropic’s research team, led by the Frontier Red Team, didn’t just identify vulnerabilities—they also developed working exploits for two of them. Here’s what makes this work groundbreaking:
- 22 vulnerabilities discovered in two weeks, surpassing monthly 2025 totals.
- 14 high-severity bugs patched in Firefox 148, including JIT miscompilation flaws.
- 90+ additional bugs found but not classified as security-critical.
- AI-generated exploits achieved arbitrary read/write access in controlled environments.
Mozilla engineers Brian Grinstead and Christian Holler confirmed that most AI-generated bug reports are typically low-quality. However, these findings stood out: each included minimal test cases, detailed proofs of concept, and candidate patches. This level of detail allowed Mozilla to verify and reproduce the issues rapidly.
Technical Insights: How AI Exploits Worked
Anthropic’s team tested whether Claude could turn vulnerabilities into functional exploits. After 4,000 API credits and hundreds of attempts, the model succeeded in two cases. One exploit targeted CVE-2026-2796, a JIT miscompilation in Firefox’s JavaScript WebAssembly engine.
The exploit leveraged a type confusion vulnerability in Function.prototype.call.bind() wrappers. By creating exploitation primitives like addrof (address leak) and fakeobj (object forgery), the AI achieved arbitrary read/write access. However, the exploit only worked in a testing environment without Firefox’s sandbox protections.
Anthropic emphasized that modern browser security features like sandboxes remain critical. The company noted:
“The exploits Claude wrote only worked in our testing environment, which intentionally removed security features like the sandbox.”
Implications for Cybersecurity
This research highlights both opportunities and risks:
Opportunities
- Accelerated vulnerability discovery: AI can scan codebases for flaws humans might miss.
- Improved patching workflows: Mozilla now integrates AI-assisted analysis into its security processes.
- Proactive defense: Anthropic’s Coordinated Vulnerability Disclosure (CVD) framework sets new standards for AI-generated reports.
Risks
- Exploit development speed: Attackers could use similar AI tools to create exploits faster than defenders can respond.
- Undiscovered bug backlogs: Grinstead and Holler suggest decades of unpatched vulnerabilities may exist in widely used software.
- Model capability gaps: While AI excels at discovery, exploit development still requires human-like reasoning in many cases.
The Future of AI in Cybersecurity
Anthropic’s work builds on earlier research where Claude identified over 500 zero-day vulnerabilities in open-source software. The company recently launched Claude Code Security, a limited research preview offering vulnerability discovery and patching tools to developers.
Mozilla’s collaboration with Anthropic has already yielded best practices for submitting AI-generated bug reports. These include:
- Providing minimal test cases.
- Detailing proofs of concept.
- Offering candidate patches.
As AI models evolve, the gap between discovery and exploitation capabilities will narrow. Anthropic warns that this progress could outpace traditional security measures within years.
What Developers and Security Teams Should Do
1. **Adopt AI-assisted security tools** like Claude Code Security for proactive vulnerability scanning.
2. **Strengthen sandboxing and isolation** to mitigate exploit risks in production environments.
3. **Collaborate with AI researchers** to refine CVD processes and improve patching workflows.
4. **Monitor AI advancements** in exploit development to stay ahead of emerging threats.
Conclusion
The discovery of 22 Firefox vulnerabilities in two weeks by an AI model signals a paradigm shift in cybersecurity. While this technology offers unprecedented speed and precision in finding flaws, it also demands a reevaluation of how we defend against threats. As Anthropic’s research shows, the future of software security will depend on balancing AI’s power with human expertise and robust defensive strategies.
Stay ahead of the curve by exploring AI-driven security tools and participating in collaborative vulnerability disclosure programs. The next frontier of cybersecurity is here—and it’s powered by artificial intelligence.
