Anthropic's Mythos model found vulnerabilities in classified U.S. government systems, official says

What happened

A U.S. official told AP that Anthropic, in partnership with American intelligence agencies, tested its most advanced Mythos model — and it found vulnerabilities in classified government computer systems.

Context and impact

The disclosure comes just weeks after the Trump administration designated Anthropic a security threat and restricted exports of Fable 5 and Mythos 5. The exercise proves frontier models are already operationally deployable for both offensive and defensive cyber tasks in a state context.

Details

• Testing conducted with U.S. intelligence agencies
• Mythos model was used
• Reported via Associated Press, covered by CNBC
• Follows export-control measures, customer lawsuits, and sovereign AI debate
• Strengthens Anthropic's argument that AI is 'critical national infrastructure'