Anthropic’s NSA red-team result helps explain U.S. model access ban

Anthropic’s Mythos model reportedly accessed “almost all” classified NSA systems within hours during a controlled security evaluation, according to comments attributed to Gen. Joshua Rudd and relayed by Sen. Mark Warner. The claim spread widely as an alleged NSA hack, but the scenario was an authorized internal red-team test using Mythos alongside defensive tools under specific simulated conditions.

The episode helps explain a June 12 U.S. government directive that barred foreign nationals, including Anthropic’s own non-citizen employees, from accessing Fable 5 and Mythos 5. Anthropic disabled the models globally, saying it could not enforce nationality-based access restrictions in practice. The move was described as the first U.S. export control aimed directly at an AI model rather than the hardware used to run it.

Anthropic says the issue involved a “potential narrow, non-universal jailbreak” that could allow Fable 5 to identify software vulnerabilities, and argues that rival models including OpenAI’s GPT-5.5 show similar behavior. The company is seeking to restore access and develop a risk-management framework with the White House, while continuing work with the NSA through Project Glasswing, where roughly six Anthropic engineers are reportedly embedded inside the agency.