The US government placed worldwide export controls on Anthropic’s Fable and Mythos models this month, forcing Anthropic to pull them down for every user on the planet, including, by some accounts, its own non-US employees. The move came days after Anthropic released Fable, a commercially restricted version of Mythos with heavy guardrails, and officials in the Trump administration became aware of a jailbreak that bypassed some of those guardrails. What exactly was demanded, by whom, and with what specificity remains opaque. What is not opaque is the outcome: a company that believed it had followed the rules found itself shut down by a government that had publicly told it to ship.
This is the central tension that Dean Ball, writing in Hyperdimensional on June 17, dissects in detail. Ball lays out the timeline carefully. David Sacks, then White House AI Czar, had dismissed Anthropic’s safety framing as a boy-who-cried-wolf problem. Emil Michael, Undersecretary of War for Research and Engineering, argued correctly that advanced cyber-capable models are an industry-wide phenomenon, not an Anthropic-specific one. Sacks praised OpenAI for releasing GPT-5.5 broadly and contrasted it favorably with Anthropic’s more cautious posture. Then, five weeks after that praise, President Trump signed Executive Order 14409 on June 2, which explicitly stated that nothing in the order should be construed to authorize mandatory licensing or preclearance for AI models.
On paper, Anthropic had every signal it needed to proceed. The EO said no mandatory licensing. Senior officials had criticized Anthropic for being too cautious. The voluntary pre-deployment testing program the EO described did not even have a deadline until July. Anthropic released Fable. The government pulled the models.
Ball’s reading, which I find persuasive, is that written policy and public rhetoric stopped mattering the moment Anthropic’s release landed inside an active political dispute. Anthropic is simultaneously designated a supply-chain risk by the Department of War, publicly aligned with Democratic candidates, and the company that has most loudly advocated for stricter AI regulation while the current administration runs on deregulation. Ball states it plainly: Anthropic is a political enemy of this administration, in part because it chose to make itself one. Releasing a powerful model under those conditions, without an explicit green light from the government, was not a policy violation. It was read as a political act.
That framing matters for builders who are not Anthropic. Ball argues that the practical reality at the frontier is now this: you need an explicit government thumbs-up before deploying a sufficiently capable model, regardless of what the written rules say. This contradicts the EO’s own text, but the EO’s text is not what governs outcomes when political relationships have broken down. The rule of law is not the rule of Washington.
The export-control mechanism used here is worth understanding. Because Anthropic cannot verify US personhood for end users at scale, a restriction on non-US persons is functionally a global takedown. That is a significant lever. The government does not need to issue a domestic ban; it can achieve the same result by invoking export-control authority, which does not require the same procedural constraints as domestic regulation.
Ball draws the right analogy to the FDA, not to argue for an FDA-for-AI, but to point out that the FDA is itself constrained by statute. It must state in writing why a drug was denied, follow defined procedures, and face judicial review. The current AI governance situation has none of that structure. The government can act on a jailbreak report with no public transparency, no written reasoning, and no appeal process. That asymmetry is the problem.
The near-term resolution Ball describes is transactional: a deal between Anthropic and the administration, not a policy framework. The medium-term fix he calls for is a technocratic statute from Congress that binds both industry and government, specifying procedures and constraints for each side. Without that, every frontier model release is a political negotiation dressed up as a compliance question.
For any lab approaching a capability level comparable to Mythos, the lesson from this week is that the voluntary pre-deployment testing program in EO 14409 is not optional in practice. The consequences of skipping it when the political relationship is strained are now documented. Plan your deployment timeline accordingly.
Analysis adapted from Dean Ball’s “Leviathan Waking,” published June 17, 2026 in Hyperdimensional.
