Anthropic's Mythos artificial intelligence model identified vulnerabilities in highly sensitive US government computer systems during a testing exercise, according to a report by the Associated Press published on Tuesday, June 23. The disclosure has drawn significant attention to both the capabilities of advanced AI and the increasingly strained relationship between Anthropic and Washington.

The tests were conducted under Project Glasswing, a restricted programme in which Anthropic partnered with US intelligence agencies to identify and fix weaknesses in critical software before malicious actors could exploit them. Senator Mark Warner of Virginia, a senior figure on national security matters in Congress, described the results at a recent congressional hearing, saying he had been told by National Security Agency chief Joshua Rudd that Mythos had "broke into almost all of our classified systems, not in weeks, but in hours." However, a US official cited by the AP offered an important clarification: while the model identified certain vulnerabilities within that timeframe, that did not mean it was able to actively exploit them in the same period — a meaningful distinction between detection and attack.

The White House, Anthropic, and the Department of Defense did not immediately respond to requests for comment. Separately, the New York Times reported that the NSA lost access to Mythos amid an ongoing dispute between the company and the US government.

The revelation comes at a turbulent moment for Anthropic, a San Francisco-based AI safety company that is reportedly preparing for an initial public offering. The firm has refused to allow the US military to deploy its models for domestic surveillance or fully autonomous weapons systems — decisions that prompted the government to place Anthropic on a national security blacklist. Washington also this month ordered the company to suspend exports of its Mythos and Fable models to all destinations worldwide and to foreign nationals, citing national security concerns.

The episode underscores a broader tension at the heart of advanced AI development: the same capabilities that make models like Mythos valuable for defensive cybersecurity work also raise profound concerns about what could happen if such tools were misused or fell into adversarial hands. The findings of Project Glasswing, however limited in public detail, are likely to intensify debate in Washington and beyond about how governments should govern, access, and constrain the most powerful AI systems.