271 vulnerabilities. That's how many critical security flaws Anthropic's Mythos model found hiding in Firefox's unreleased code—a 12x improvement over its predecessor's haul of 22.
The numbers, published Tuesday by Mozilla in a blog post, mark the first time an AI security model has demonstrated production-scale effectiveness on a major real-world codebase. In Firefox 150, Mythos Preview identified 271 security-sensitive bugs. Last month, Anthropic's Opus 4.6 managed only 22 when reviewing Firefox 148. The leap from two dozen findings to nearly three hundred represents something qualitatively different: not incremental progress, but a step-change in automated vulnerability discovery.
Firefox CTO Bobby Holley didn't hedge. "Defenders finally have a chance to win, decisively," he wrote, framing the results as a turning point in the decades-long asymmetry that has favored attackers. The traditional calculus in cybersecurity has been brutal: defenders must find and fix every flaw, while attackers need only exploit one. AI that can scale bug discovery could shift that balance.
But the Firefox team's own assessment is more measured. AI won't transform cybersecurity overnight, they warned. Developers face a challenging transition period as the technology matures and both defenders and attackers learn to wield increasingly capable AI tools. The race isn't over—it's entering a new phase with uncertain outcomes.
The significance of this case study extends beyond Firefox. Mozilla's codebase is massive, maintained by a global distributed team, and has been a target for nation-state hackers and criminal syndicates alike. If Mythos can find hundreds of previously undetected vulnerabilities in that environment, the implications for software supply chains everywhere are substantial. Every major platform, from banking infrastructure to critical utilities, runs on codebases with similar blind spots.
The debate between optimism and caution reflects genuine uncertainty about AI's long-term trajectory in security. If defenders gain AI-powered bug hunting, so do attackers. The same model capabilities that found 271 Firefox flaws could theoretically help adversaries identify zero-days. Anthropic's decision to restrict Mythos Preview to "critical industry partners" reflects this tension. Open access could accelerate both defensive and offensive applications simultaneously.
What the Mozilla data actually demonstrates is narrower but more valuable than the hype: AI security models can now work at scale in production codebases. The transition will be uneven, the dual-use risks real. But for the first time, there's concrete evidence that defenders have a tool capable of matching the speed and scope of modern cyberattacks. The question is no longer whether AI will reshape security—it's whether that reshaping benefits those who move fastest.