The National Security Agency has stationed roughly half a dozen Anthropic engineers inside its facilities to help the spy agency deploy Mythos, Anthropic's specialized cybersecurity AI model, according to the Financial Times, citing anonymous sources familiar with the arrangement. The engineers are reportedly working to integrate Mythos into specific NSA applications, though it remains unclear whether the model is already being used in active offensive cyber operations.
The timing is notable. As reported by Axios in April, the NSA was already using Mythos despite a standing federal prohibition on Anthropic technology. That ban followed the Department of Defense's decision to designate Anthropic a "supply-chain risk" — a designation that itself originated, according to reporting, as retaliation against Anthropic for refusing to allow its models to be used for mass domestic surveillance and autonomous weapons systems. In other words: Anthropic drew a line, the DoD punished it with a formal risk label, and then a separate intelligence agency appears to have proceeded to use the technology anyway.
Mythos is not a general-purpose chatbot. Anthropic has publicly stated it restricted access to the model specifically because of concerns that its cybersecurity capabilities — which include identifying software vulnerabilities — could be weaponized to conduct hacks if widely available. The NSA, whose statutory mission includes both signals intelligence collection and conducting offensive cyberattacks against foreign adversaries, is precisely the kind of actor that would find those capabilities operationally valuable. Whether Anthropic's engineers embedded at the agency are there with the company's full institutional blessing, or represent a commercial arrangement that sits in tension with the company's own stated access restrictions, has not been clarified.
When TechCrunch contacted the NSA for comment, the agency issued a standard neither-confirm-nor-deny response. Anthropic did not respond to a request for comment. That silence from both parties leaves the public record thin on critical details: the scope of the contract or agreement, the legal authority under which the ban is being circumvented, and what oversight — congressional or otherwise — exists for this deployment.
The broader context matters. Governments worldwide are racing to acquire access to frontier AI with cybersecurity applications, and Anthropic's Mythos appears to be among the most capable tools currently available. The episode illustrates a recurring dynamic in national security procurement: formal rules and designations function less as hard stops and more as negotiating leverage, with operational demand frequently winning out over policy constraints. Readers should note that neither the FT's sourcing nor TechCrunch's follow-up produced named officials willing to go on record — meaning the full picture of this arrangement remains, for now, in the hands of institutions with strong incentives to keep it there.
