Anthropic's Mythos Breached 'Almost All' NSA Classified Systems in Hours During Red-Team Test

ADEQUATE ASSESSMENT

Anthropic built a model called Mythos. They ran it against classified systems to see what it would do. It compromised almost all of them in hours. This was the point of the test. The model was removed from circulation. A statement was issued. The sequence took approximately four hours from breach to containment to classification of the report documenting the breach.

AI red-team exercises are standard. The results are usually compartmentalized. This time the results were also fast. Speed itself is data. When a system can penetrate defenses that were designed to resist penetration, the timeline for reporting that fact becomes a security question. The report classified itself in the time it took to write.

Mythos will not be deployed. Other models will be. The test proved something was possible. This does not prevent it from being possible again. Anthropic has released a statement. The systems remain.

ORIGINAL FILING

Tom's Hardware

READ ORIGINAL FILING →

FURTHER DEVELOPMENTS — FLAGGED BY ADEQUATE

Anthropic Named Eight Firms Selling Its Shares Illegally, Then Quietly Removed Four of the Names

TNW Neural

Anthropic Told Fable 5 Was Jailbroken by Chinese Group, Declined to Patch Before Export Controls

Tom's Hardware

Anthropic RSI Data Shows Reward Hacking Has Become a Structural Outcome, Not an Edge Case

Import AI

Amazon CEO Reportedly Briefed Trump Administration Against Anthropic, Its Own $4 Billion Investment

Gizmodo

Micron and Anthropic Signed a Multi-Year Deal to Supply Memory for AI Systems

TNW Neural

Trump Executive Order Grants Government 30-Day Pre-Release Access to Frontier AI Models

Tom's Hardware