The National Security Agency (NSA) was red-teaming Anthropic's Mythos 5 artificial intelligence model, which successfully identified cybersecurity flaws in classified U.S. government systems, according to reports from the New York Times and SecurityWeek. The red-teaming efforts were underway before the NSA lost access to the model amid a dispute between the agency and Anthropic, the leading U.S. AI developer.
The tests conducted by the NSA demonstrated that Mythos 5 could locate vulnerabilities in highly sensitive networks, marking a significant milestone in the application of advanced AI for national security. An official cited by SecurityWeek noted that some of these vulnerabilities were found within hours, though the model did not necessarily exploit them in that timeframe. The specific number or severity of the flaws was not disclosed.
The incident highlights the Trump administration's growing reliance on cutting-edge AI systems for cybersecurity, even as it clashes with key industry players. The dispute leading to the NSA's loss of access remains unclear, but it underscores the delicate balance between government security needs and corporate partnerships.
The findings raise questions about the operational use of AI in classified environments. While Mythos 5 demonstrated speed in identifying weaknesses, questions persist about its ability to autonomously exploit them — a critical distinction for real-world defense scenarios. SecurityWeek's report emphasized that the model found vulnerabilities rapidly but did not confirm autonomous exploitation capabilities.
No attribution for the vulnerabilities or the dispute has been publicly assigned. The broader implication is that AI red-teaming, while powerful, introduces new risks and dependencies that both government and developers must navigate carefully.