Capability based controls on frontier AI, the same rules designed to keep models like Anthropic's flagship Mythos 5 and Fable 5 frontier models out of adversary hands, have now cut US intelligence agencies off from those very models.
The instrument that produced this outcome was not designed for it. The US export-control regime for frontier AI gates on the technical properties of a model, not on the identity or intent of the buyer. Once a model crosses the "exceptionally capable" threshold, the same legal trigger fires for an adversary in Beijing and for the NSA in Fort Meade. That is the design. It is now producing visible costs inside the US government.
The proximate form of those costs: the NSA was among a limited group granted early access to Anthropic's Mythos 5 and Fable 5 for cybersecurity work. Anthropic withdrew that access after the Commerce Department recently imposed export-control measures, framed on national-security grounds. NextGov framed the same event as a model suspension triggered by an export-control order. The New York Times report that surfaced Monday via Firstpost put the operational loss on the record.
What the wire coverage did not say clearly, until the NSA Director's office clarified this week, is what those models were actually being used for. At a recent hearing, Senator Mark Warner, the senior Democrat on the Senate Intelligence Committee, relayed a remark from NSA Director Gen. Joshua Rudd: Mythos, the general said, "broke into almost all of our classified systems, not in weeks, but in hours". Trade press read this as an autonomous breach. Security Affairs and Tom's Hardware carried the framing. The NSA Director's statement to the Senate Intelligence Committee made clear the remark described NSA cybersecurity teams using Mythos inside controlled red-team exercises—not an autonomous compromise of classified networks.
That distinction is not a walkback. It is the operative fact for the policy question. A frontier model used by US defenders to find weaknesses is a procurement asset. The same model in adversary hands is a national-security liability. Capability-based controls, as currently written, cannot tell the two apart. That is the design choice the present dispute is exposing.
The asymmetry is what makes the suspension unusual. Washington has spent the last year publicly championing frontier AI as essential to US cyber and intelligence work, even as it tightened the export controls meant to keep that same class of model out of adversary hands. The NSA's loss of Mythos 5 and Fable 5 is the first visible case where the two postures collided in the same agency, at the same model, in the same month.
The pattern has precedents, even if the current case is the sharpest. In March 2026, Anthropic was designated a supply-chain risk under a separate review, which has shaped how government contractors handle Anthropic tooling. That designation is contextual rather than proximate for the current suspension, but it illustrates the same structure: a regulator draws a line around the supplier, and the line carries consequences the regulator did not intend for its own principals. Volkov Law has tracked the broader collision for AI governance.
The open question is whether the regime can be recalibrated. A model-level capability threshold treats every "exceptionally capable" model as equally restricted, regardless of buyer. A user-level regime, by contrast, would license access by buyer and end-use, closer to how the Commerce Department already handles other dual-use technologies. The current dispute, read structurally, is the first test of which model the United States wants. The intelligence community's vote, expressed through the operational loss of two frontier models it was actively using, is now on the record.