A United States government official has confirmed to the Associated Press that Anthropic’s Mythos artificial intelligence model identified vulnerabilities in highly sensitive and classified US government computer systems during a controlled testing exercise, with the model locating certain weaknesses within hours rather than the days or weeks that conventional security testing might require. The official, who spoke on the condition of anonymity given the sensitive nature of the subject matter, clarified that while the model identified the vulnerabilities within that compressed timeframe, this did not necessarily mean it was able to actively exploit them within the same period. The disclosure represents one of the most significant public confirmations to date of the capability gap that frontier AI models are opening in the domain of offensive cybersecurity, and arrives at a moment of intense scrutiny over how governments and technology companies should manage the proliferation of AI systems with capabilities that touch directly on national security.
The testing was conducted through an Anthropic initiative called Project Glasswing, which brought together technology companies and other industry partners with the explicit goal of securing the world’s critical software infrastructure against the severe risks that the Mythos model could pose to public safety, national security, and economic stability. The programme reflects a growing recognition within both the technology industry and the US intelligence community that the most capable AI models must be stress-tested against real-world critical systems before they are widely deployed, and that the organisations developing these systems bear a direct responsibility for understanding and mitigating their potential for harm. The involvement of US intelligence agencies in the testing exercise underscores the degree to which frontier AI capability has become a matter of national security concern rather than simply a commercial or academic question.
The findings had been briefly referenced in a Senate hearing on 11 June before the Senate Committee on Banking, Housing, and Urban Affairs, where it was disclosed that the tool had broken into almost all classified systems tested, not in weeks but in hours. The National Security Agency declined to comment when contacted, and an Anthropic spokesman also declined to provide a response, leaving the official who spoke to the Associated Press as the sole confirmed source for the specific details of what the testing revealed. The disclosure lands in the context of significant recent turbulence surrounding Anthropic’s most powerful models, with the company having suspended worldwide access to both Mythos and its companion model Fable 5 on 13 June following a US export control directive tied to security concerns, a decision that cut off Australian and other international users without prior warning and that has drawn fresh attention to the question of how export controls and national security considerations will increasingly shape global access to frontier AI systems.