Anthropic's Mythos AI Model Finds Vulnerabilities in US Classified Systems

Anthropic’s advanced language model, Mythos, was used in a red‑team exercise with U.S. intelligence agencies under the Project Glasswing initiative. In the test the model identified security flaws in highly classified government networks within hours, a speed highlighted by Senator Mark Warner who quoted NSA and Cyber Command chief Joshua Rudd as saying the tool “broke into almost all of our classified systems, not in weeks but in hours.” The agency stressed that finding a vulnerability does not mean the model could exploit it in the same timeframe.

Following the disclosure, the Trump administration issued export‑control directives that barred foreign nationals from accessing Anthropic’s latest models, Mythos 5 and Fable 5, and Anthropic temporarily disabled them for all customers. The move prompted a backlash from cybersecurity firms and experts who argued the models are valuable for defensive work and that the restriction could benefit adversaries. Meanwhile, a limited rollout of Mythos to selected organizations, including the Australian government, continues under tighter controls.

The episode has spurred legislative proposals, such as the Combat Emerging Threats to Critical Infrastructure Act, aimed at overseeing frontier AI systems before public release, and has intensified the debate over AI’s role in national security.