How Mozilla uses GPT-6 Cyber to find and fix Firefox vulnerabilities

Categories: AI, Product

Summary

Mozilla deployed an agentic AI security pipeline that found bugs GPT-6 Cyber missed in static analysis alone, preventing zero-days at PWN2OWN 2026—proving that different AI models discover different vulnerabilities even in repeatedly scanned codebases.

Key Takeaways

  1. Implement LLM-based file prioritization scoring using domain-specific threat models rather than alphabetical scanning; Mozilla weighted files by memory safety risk, exposure to web content, and historical scan frequency to guide agent focus.
  2. Design agentic security loops with deterministic verification—agents test bug hypotheses through dynamic fuzzing tools in isolated VMs, iterate rapidly on false theories, and only surface findings with proven exploitability.
  3. Embed decades of institutional security knowledge (bug bounty programs, fuzzer infrastructure, debuggable builds, technical documentation) into AI tooling; existing red team primitives become force multipliers for agents.
  4. Diversify vulnerability detection across multiple AI models—OpenAI Cyber found distinct bugs that competing researcher tools missed, indicating model-specific detection blind spots require parallel scanning strategies.
  5. Establish rapid patch velocity workflows before deploying agentic scanning; Mozilla's ability to fix, uplift, and ship fixes quickly before PWN2OWN turned discovered vulnerabilities into defensive advantages rather than liabilities.

Related topics

Transcript Excerpt

Well, >> hey Brian, thanks for stopping by OpenAI. >> Happy to be here. >> There was recently a pone to own competition. Could you tell me a little bit about how that went for you all and also like what is pone to own? >> So pone to own is an annual uh hacking competition and browsers are one of the top targets. Of course, every year we use that as an opportunity sort of as a fire drill for what it would be like if there were a real security critical bug in the wild that we knew to be exploited because we know that researchers are going to show up and attack Firefox. And it's a great controlled environment for us to to practice that. And we knew coming into it this year in 2026 that many researchers are using these agentic scanning tools and there's a lot of bugs to find. Earlier in the ye…

More from OpenAI