A 'supply chain risk' label and an export ban on its flagship AI model: why Washington is treating one frontier lab differently from its closest rival, weeks before a $965 billion IPO.
The Friday before last was the kind of coincidence that crystallizes a pattern. The Commerce Department told Anthropic it would relax export controls on the company's flagship artificial intelligence model, Mythos, allowing limited release to some companies and government agencies after a roughly two-week ban that had forced Anthropic to pull the model from every user. On that same day, OpenAI announced it was voluntarily withholding wide release of its newest model, identified in press coverage as GPT-5.6, at the US government's request, offering only a staged preview to trusted partners. Both companies had new frontier models ready. Only one was allowed to ship them widely. Fortune's reporting on the divergence and the Guardian's account of OpenAI's release decision treat the two events as separate stories. Read together, they describe a two-track system Washington is now running for frontier artificial intelligence labs.
To understand what happened to Mythos, it helps to know what the model is and what the "supply chain risk" label that the Pentagon hung on Anthropic actually does. Mythos is Anthropic's frontier large language model, the company's most capable system, with cyber-offensive capabilities that go beyond consumer products. Fable is a commercially released variant, broader but less powerful, the version most enterprise customers would actually run. A "supply chain risk" designation is a Pentagon tool borrowed from the defense-procurement world: it lets the Department of War treat a company's technology as if it could compromise the integrity of military supply chains, justifying restrictions on contracts and information-sharing that would otherwise be unavailable. A March 2026 Mayer Brown explainer walks through the legal mechanics, and Fortune reported the Pentagon's formal designation that month.
The designation followed Anthropic's refusal of specific Pentagon contract language. The company declined to accept the conditions the Defense Department required as a condition of its work, and the "supply chain risk" label followed in April 2026. About two weeks before Fortune's publication, the Commerce Department's Bureau of Industry and Security took a more aggressive step: it imposed export controls on Mythos itself, and on the commercial Fable variant, after researchers discovered a Fable jailbreak that could let users reach Mythos's full cyber capabilities. Anthropic was forced to disable Mythos for all customers. CNBC reported the export action; Anthropic's own account appeared on X, and Just Security's legal analysis argued that blocking individual model weights was without clear precedent. Tech Policy Press reached the same conclusion independently.
The mechanism matters more than the personalities. Export controls have historically been applied to whole companies or whole categories of technology, not to a single model's weights. Using them on Mythos turns the lever from "this firm cannot sell to certain buyers" into "this capability cannot leave the building until the government says otherwise." That is a different tool, and Just Security's analysis argues it has no obvious prior use. It also turns compliance into a precondition for release: a frontier lab can either pre-coordinate with Washington or watch its most important product get frozen.
OpenAI's release on the same Friday illustrates the alternative path. Forbes reported in May 2026 that the Center for AI Standards and Innovation had amended partnerships with Google DeepMind, Microsoft, and xAI, allowing pre-release government evaluation of new models, the same arrangement OpenAI and Anthropic operated under during the Biden administration two years ago. OpenAI's decision to withhold wide release of its latest model, identified in coverage as GPT-5.6, at the government's request fits that pattern. Anthropic, by contrast, is no longer in the pre-release evaluation arrangement, having declined the contract terms that would have kept it there.
The political backdrop is loud but secondary. President Trump posted that Anthropic consisted of "leftwing nut jobs" trying to "strong-arm the Department of War." Emil Michael, the Under Secretary of Defense for Research and Engineering, called CEO Dario Amodei "a liar" with a "God-complex" at public events. Defense Secretary Pete Hegseth called Amodei "an ideological lunatic" at an April Congressional hearing. David Sacks, Trump's former AI and crypto czar, has accused Anthropic of running a "sophisticated regulatory capture" strategy and supporting state-level AI laws to "backdoor Woke AI." The attacks are on the record in Fortune's reporting. What they describe is not new friction between a White House and a tech CEO; they are the temperature of a dispute whose substantive content is a procurement contract, an export-control regime, and a question about who decides when a frontier model ships.
The stakes are unusually concrete. Fortune reports Anthropic at a $965 billion valuation, with IPO paperwork filed in recent months expected to become public soon. Anthropic's own public framing, posted to X after the export controls, called the episode potentially existential, unusual language for a company that wants public-market investors to underwrite its growth.
The next test is whether the Mythos precedent becomes the template. The Friday relaxation of the Mythos controls was partial and reversible; the legal mechanism that produced the ban is now on the books and available to any future administration that wants to use it. For Anthropic, that question lands in an S-1 filing. For the rest of the frontier labs, it lands in the contract language they will be asked to sign next.