June 23, 2026

Anthropic’s Mythos AI breaches NSA systems during red-team test

Anthropic’s Mythos AI reportedly breached “almost all” classified systems of the NSA during a controlled red-team test, prompting the U.S. government’s swift ban on its flagship models. This revelation surfaced when Sen. Mark Warner disclosed that Gen. Joshua Rudd, head of the NSA and U.S. Cyber Command, informed him about the breach occurring within hours, a claim that went viral despite later clarifications that the evaluation was authorized and conducted under specific simulated conditions. The June 12 ban, which notably marked the first application of export controls directly to an AI model rather than its hardware, raised significant national security concerns about vulnerabilities. Following this, Anthropic, which has engineers embedded within the NSA as part of its ongoing collaboration through the Project Glasswing program, stated that the flagged breach involved only recognized bugs and is currently working on restoring access to its models while preparing a risk-management framework with the White House.

Anthropic: Anthropic is an AI company focused on developing advanced frontier models including Mythos and Fable series for research and enterprise use. In this news, its Mythos model underwent a controlled red-team security evaluation that revealed rapid access to classified systems, prompting a U.S. government ban on foreign access and leading the company to disable the models globally while pursuing a risk-management framework with the White House.
Joshua Rudd: Joshua Rudd is a U.S. General serving as head of the NSA and U.S. Cyber Command. In this news, he provided the briefing to Senator Warner detailing the Mythos model’s rapid breach of NSA systems during an authorized internal evaluation, supplying key context for the government’s subsequent actions.
Mark Warner: Mark Warner is a U.S. Senator serving as vice chair of the Senate Intelligence Committee. In this news, he publicly referenced a briefing from NSA leadership on the Mythos model’s performance during the red-team test, contributing to the viral discussion and clarification around the reported breach.
National Security Agency: The National Security Agency is the primary U.S. intelligence agency responsible for signals intelligence, cybersecurity, and protecting classified networks. In this news, the NSA conducted an internal red-team test of Anthropic’s Mythos model that demonstrated its ability to breach classified systems within hours, which informed the subsequent government directive restricting access to the AI models.

AI Security Evaluation: Controlled red-team testing of frontier AI models can reveal rapid access capabilities to sensitive systems under simulated conditions.
Government AI Oversight: The U.S. government applied export controls directly to an AI model for the first time, citing national security concerns over potential vulnerabilities.
Public-Private AI Collaboration: Anthropic continues specialized work with the NSA through its Project Glasswing program, embedding engineers for operational adaptations of models like Mythos.

AI tools built by Emerald Force

Built and supported by Emerald Force.

You might also like