← Back to feed

Parts of the National Security Agency have lost operational access to Anthropic’s Mythos 5 model following Trump administration export restrictions that took effect after the Mythos shutdown in June. The loss was confirmed by two people familiar with the matter and reported by Defense One and the New York Times.

The loss is operationally significant. NSA had been using Mythos for cyber work — vulnerability discovery, code analysis, and offensive and defensive security operations. The agency’s use of Mythos was disclosed earlier this year, when NSA was simultaneously deploying Mythos through Anthropic’s Glasswing coalition while NSA’s parent department, the Department of Defense, had identified Anthropic as a potential AI supply chain risk.

That contradiction has now resolved in the direction of disruption.

The Split Access Problem

Not all users lost access equally. Early commercial and government testers who had received access to a Mythos preview before the export order were permitted to retain that access under a grandfather arrangement. Reuters reported this carve-out in mid-June, framing it as a product continuity provision for organisations that had begun integrating Mythos into production workflows.

NSA’s broader operational deployment fell outside that window. The agency had scaled beyond the early-tester tier, which meant it was subject to the full export restriction when the White House order came through. The result is a two-tier access structure: a small group of early adopters still running Mythos, and the main body of institutional users — including parts of the US intelligence community — locked out.

The Timing

The Mythos shutdown was triggered in June after security incidents exposed the model’s access prior to official release, and after the White House moved to restrict export of frontier AI models that could be used for offensive cyber operations. Anthropic was given limited options: maintain restricted access for a small list of approved entities, or face broader government intervention.

The 90-institution expansion Anthropic had sought — to move Mythos access from roughly 50 organisations to over 120 — was blocked by the White House before the June shutdown. That blocked expansion included several more intelligence and defence agencies that had been queued for access.

Semafor reported in June that a China-linked group had accessed Mythos prior to the export order, which became part of the justification for the restrictions. NSA was cited in earlier reporting as both a user of Mythos for cyber offense capability and simultaneously a subject of concern at the intelligence community level regarding foreign access to the same tools.

The Structural Issue

The episode exposes a gap in how frontier AI models are treated as intelligence infrastructure. NSA built operational workflows around Mythos. When export restrictions landed, those workflows stopped. The agency, per sources, has limited alternative tools that match Mythos’s capability in the specific tasks it was being used for.

The White House is reportedly building a jailbreak-resistance scoring benchmark as a condition for restoring Mythos access more broadly. Until that process resolves, the operational gap at NSA and other restricted users remains open. No return timeline has been made public.