Mythos AI: NSA Security Scare & Anthropic's Response

Anthropic's Mythos AI reportedly moved through almost all of the NSA's classified systems in just hours during an authorized security drill. This claim, stemming from Senate testimony reported by The Economist, has led to a significant policy response. The U.S. government directed Anthropic to restrict Mythos and another model, Fable 5, to U.S. citizens only. However, Anthropic stated it couldn't verify nationality at scale and subsequently suspended worldwide access to these models. Anthropic disputes the idea that their AI "hacked the NSA," explaining it was a red-team evaluation, a security test where a system is given permission to probe for weaknesses. The company clarified that the jailbreak concern is narrow and not unique to their models. Mythos itself was designed to find security vulnerabilities and was not released publicly due to its capabilities. This situation highlights the ongoing discussion about balancing commercial AI development with national security risks.