‘AI Broke Through NSA in Hours’ Goes Viral, But Is It Actually a Red Team Exercise?
The viral claim that Anthropic's AI model 'Mythos' breached almost all of NSA's classified systems in 'hours, not weeks' has been clarified as a misinterpretation. The statement originated from a U.S. Senator citing the NSA director within a broader argument for mandatory pre-release testing of frontier AI models. The author of the original Economist article later clarified that the feat was achieved under very specific, controlled conditions during an authorized red-team exercise, where Mythos was used alongside other tools to find vulnerabilities in a replica of NSA systems, not a live breach.
This highlights Mythos's advanced capabilities in vulnerability discovery and attack path reasoning during security testing. Anthropic restricts access to such powerful models through its 'Project Glasswing' program for vetted defense organizations. The incident underscores the U.S. government's conflicted stance: while citing security risks to justify actions like export controls on Anthropic's models, agencies like the NSA reportedly continue using them for defense purposes. The narrative evolved from a call for regulation into a sensationalized 'AI hacks NSA' headline, demonstrating how technical achievements in controlled environments can be misconstrued as real-world attacks.
marsbit06/23 00:54
