Anthropic’s Mythos model identified vulnerabilities in highly sensitive, classified US government computer systems within hours during a restricted testing programme, a finding that immediately hardened Washington’s stance on controlling frontier AI access. The exercise, conducted under Project Glasswing in collaboration with US intelligence agencies, was designed to find and fix critical software weaknesses before adversaries could exploit them. Senator Mark Warner, citing the NSA chief, stated the model “broke into almost all of our classified systems, not in weeks, but in hours,” though unnamed officials later clarified that Mythos detected the weaknesses but did not exploit them in that timeframe.

The capability demonstrated by Mythos—rapid vulnerability discovery without active exploitation—arrived amid already strained relations between Anthropic and the Trump administration. The company had previously refused to permit military use of its AI for domestic surveillance or fully autonomous weapons, leading the government to place it on a national security blacklist. Within days of the testing disclosure, the Commerce Department ordered Anthropic to disable its latest Fable 5 and Mythos 5 models for all foreign nationals, citing cybersecurity fears. Anthropic complied by cutting off all customers globally, a move it called unnecessary, while the administration reportedly gave the firm just 90 minutes to act.

The disruption has rippled across borders and industries. A US legal technology firm, Legion LegalTech, sued the government, arguing the directive unlawfully severed access for its Canada-based development team and caused immediate, irreparable harm. In Canada, Cohere CEO Aidan Gomez warned that the episode exposes the structural risk of renting AI from foreign powers, urging democratic nations to pursue sovereign AI capabilities or form trusted alliances to control the full supply chain. More than 100 cybersecurity executives, including leaders from Adobe and Nvidia, urged the administration to lift the restrictions, contending that sidelining Anthropic’s models could weaken US defences and that Mythos is effective but not uniquely so compared to other foundation models.

The legal challenge in Washington, D.C., federal court seeks to vacate the directive, while Anthropic has signalled willingness to cooperate more closely with the White House to resolve the standoff. The company plans to restore the models to its subscription offerings once sufficient compute capacity is available, but the timeline hinges on regulatory clearance. The broader policy trajectory is set by a presidential executive order requiring federal vetting of advanced AI systems for national security risks before public release, making the outcome of this dispute a potential precedent for future AI export controls.