Anthropic’s testing reveals classified U.S. government systems have vulnerabilities, says anonymous source

Within hours of testing, Anthropic’s Mythos model discovered security vulnerabilities in U.S. government computer systems. This was part of a collaboration with U.S. intelligence agencies under Project Glasswing, aimed at securing critical software. Democratic Sen. Mark Warner highlighted the model’s rapid success in identifying these issues during a Senate hearing. Despite this cooperation, tensions exist between Anthropic and the Trump administration, particularly over the military’s use of AI and restrictions on Anthropic’s models. A directive now limits foreign access to Anthropic’s AI, following an executive order for vetting AI systems for national security risks. Anthropic complied by disabling its models, though it questioned the necessity of the government’s actions. Cybersecurity experts have urged the administration to reconsider the directive, fearing it may benefit adversaries. QUESTION: How might the discovery of vulnerabilities by AI models like Mythos influence future cybersecurity measures and policies?