The trigger was a narrow jailbreak (a prompt trick that bypasses safety guardrails) — a feat a rival top tier AI model can already produce without one. The lever was an export restriction, a trade policy tool used to compel the takedown.
The United States government forced Anthropic to take two of its frontier AI models, Fable and Mythos, offline on Friday evening, not because they malfunctioned, but because a narrow prompt trick let someone elicit outputs that a rival frontier model already produces without any trick at all. According to Mowshowitz's reporting, the lever Washington pulled was not a court order, a new rule, or a congressional vote. It was an export restriction, a trade-policy tool repurposed, in real time, as an AI enforcement mechanism.
That distinction is the story. The trigger was a jailbreak, a method of crafting inputs that bypasses a model's stated safety guardrails. The takedown was not. According to analyst Zvi Mowshowitz's account on his Substack, the White House first asked Anthropic to pull Fable and Mythos. Anthropic's chief executive, Dario Amodei, reportedly pushed back, arguing no fix was needed. The White House then hit the company with an export restriction it knew would force the models offline for every user (The Once And Future Fable #2).
Mowshowitz's narrative carries an explicit "fog of war" qualifier. He is the sole named source, and the key claims about White House intent and Amodei's stance are his reporting and interpretation, not on-the-record statements from either party. The capability comparison, that OpenAI's GPT-5.5 produces the same class of outputs without a jailbreak, is also his assertion, not an independent demonstration. Treating the post as a primary news item would overstate what is actually known.
What the post does clearly describe, however, is a mechanism. An export restriction, the kind of tool normally used to keep advanced chips or sensitive technology out of foreign hands, was deployed against a domestic AI lab to compel a model takedown. No public hearing, no rulemaking, no judicial review. A trade-policy lever, in Mowshowitz's telling, was used as informal AI governance.
The precedent matters more than the specific models. The next time a frontier system produces a politically uncomfortable output, the path of least resistance is now a phone call, followed by a trade restriction if the lab hesitates. Engineers can be overruled by officials who, on the evidence in Mowshowitz's account, did not accept the technical argument that the demonstrated outputs were already in the wild on a competing model.
The criticism is not new. Dean W. Ball, a commentator frequently cited in AI policy debates, was quoted by Mowshowitz as saying that, "in a real sense the government just made my world dumber." The line captures the substantive grievance: a tool that worked for many users is gone, not because it failed, but because someone with authority to restrict its export decided it should be.
What to watch next is whether the export-control mechanism becomes a recurring instrument, used quietly each time a model produces something Washington finds inconvenient, or whether the precedent draws enough public attention to force a conversation about which lever is appropriate. Fable and Mythos are not the first models to come down under pressure. They may be the first to come down under this particular kind of pressure.