Anthropic said it is restricting access to its Claude Mythos Preview model after it uncovered thousands of critical vulnerabilities across operating systems, browsers and core software infrastructure.
The company said the model’s coding capabilities now rival top human experts, enabling it to identify high-severity flaws across major systems, including long-standing bugs.
“Given the rate of AI progress, it will not be long before such capabilities proliferate, potentially beyond actors who are committed to deploying them safely.”
Anthropic launched Project Glasswing alongside more than 40 partners including Amazon Web Services, Apple, Google and Microsoft to use the model defensively and patch vulnerabilities before they are exploited.
The model has already identified decades-old issues, including a 27-year-old bug in OpenBSD and vulnerabilities in widely used systems such as Linux, FreeBSD and cryptographic protocols.
Anthropic said 99% of the vulnerabilities found remain unpatched, underscoring the scale of the cybersecurity challenge as AI capabilities rapidly advance.
While the firm expects AI to ultimately strengthen global cybersecurity, it warned the transition period will be risky as both defenders and attackers gain access to increasingly powerful tools.