Anthropic's Claude Mythos Finds Thousands of Zero-Days Across Every Major OS and Browser
Anthropic's unreleased Claude Mythos Preview model—shared under Project Glasswing with ~40 organizations—has identified thousands of zero-day vulnerabilities across every major operating system and browser, raising urgent questions about AI's dual role in cybersecurity.
Original sourceAnthropic has quietly distributed an early version of its most powerful unreleased model, Claude Mythos Preview, to roughly four dozen organizations under a program called **Project Glasswing**. The list includes major tech companies, banks, and government-adjacent research institutions. The model's stated purpose: find and help patch zero-day vulnerabilities before adversaries can exploit them.
The results have been alarming in both directions. Claude Mythos Preview has reportedly identified thousands of previously unknown vulnerabilities across every major operating system—Windows, macOS, Linux—and every major web browser. The Air Street Press State of AI report confirms that frontier models have "crossed into offensive cyber operations," with Mythos clearing a 32-step network penetration simulation in 3 of 10 attempts. OpenAI's GPT-5.5 reportedly replicated comparable results weeks later.
Project Glasswing frames this capability as defensive: share Mythos Preview with trusted organizations, use it to find and patch vulnerabilities before bad actors do. But cybersecurity researchers, including Bruce Schneier, have raised pointed questions about the dual-use nature of the release. A model that can autonomously find and chain zero-days is by definition also a model that can exploit them—and the decision about who gets access is Anthropic's alone, without regulatory oversight.
The UK's AI Security Institute has stated that "frontier cyber-offence capability is doubling every four months." Glasswing is Anthropic's attempt to stay ahead of that curve through coordinated disclosure at scale. Whether it succeeds—and whether the model stays out of adversarial hands—will define a significant chapter in AI governance over the coming year.
Panel Takes
The Builder
Developer Perspective
“If Mythos can actually find thousands of zero-days across major OSes, responsible disclosure at that scale would meaningfully improve global software security. The engineering challenge of patching at that volume is enormous, but the alternative—leaving the bugs in—is worse.”
The Skeptic
Reality Check
“A private company deciding who gets access to a model capable of autonomous cyberattacks is not a governance framework—it's a liability waiver. One leaked Mythos API key or one compromised partner organization is a serious global security incident. The regulatory vacuum here is dangerous.”
The Futurist
Big Picture
“We're watching the first real proof-of-concept for AI as a force multiplier in offensive and defensive cyber at national-infrastructure scale. Glasswing may be remembered as the moment AI security capabilities became undeniably real—for better and for worse.”