The government switched off the world's best AI model over a bug it won't show us

Timeline, because precision matters here. 9 June: Anthropic launches Fable 5. The model is publicly available, generating users, being praised. 12 June, evening: a directive lands — reportedly a letter from Howard Lutnick at Commerce to Dario Amodei — barring any foreign national from access. Not a subset of foreign nationals. Not nationals of adversary states. Any foreign national. By Friday the model is dark: Fable 5 and the restricted Mythos 5, switched off. Not throttled, not geo-blocked by country. Off. For everyone. Because once you ban 'every foreign national', you've effectively banned global access, including Anthropic's own non-American staff. The only way to comply is a full kill.

Here is where the story gets uncomfortable. Anthropic's public account is specific: it received verbal notice — not a formal technical brief, not a written finding, verbal — of a 'potential narrow, non-universal jailbreak.' The company then watched a demo. That demo, per Anthropic, turned up 'a few previously-known minor vulnerabilities.' Not a novel zero-day exploit. Known minor bugs. Anthropic reviewed the evidence, concluded it does not justify pulling the model, and said so publicly. The government's actual technical reasoning? Not published. We are asked to trust a kill order on a product that was live and in active use, based on a vulnerability assessment that exists — as far as the public record goes — as a verbal conversation and a contested demo.

And then there is the kicker. Anthropic's own rebuttal names OpenAI's GPT-5.5 as having the same underlying capability. Same capability. Different treatment. As of the date of this report, GPT-5.5 is running normally. No directive. No suspension. No verbal jailbreak notice. So either the risk is real and applies to both models — in which case the asymmetry is indefensible and the government has a serious explanation gap — or the risk is narrow enough that restricting one company's models is disproportionate, in which case the recall is worse than theatre. There is no reading of the facts that makes both choices coherent.

The asymmetry demands an explanation. None has been offered.

Anthropic is weeks from a major IPO. That is not a trivial detail. A company approaching public markets has every commercial incentive to resolve a government dispute quickly and quietly — to demonstrate it can work with regulators, to avoid the narrative of an AI lab fighting Washington, to keep the S-1 filing clean. The pressure to settle is structural, not a conspiracy. And a quiet settlement, once it happens, will make it much harder to press for the public accounting that this episode deserves. The government will have gotten what it wanted without ever explaining why; Anthropic will have gotten its models back and will have little appetite to re-litigate it.

That window — before the quiet resolution closes it — is exactly when someone needs to be asking: what was the actual vulnerability? Why verbal notice only? Why is GPT-5.5 unaffected? Does the export-control framework have any public-accountability mechanism when applied to a deployed consumer product? These are not hostile questions. They are the minimum a democratic society should require before accepting 'national security' as the sufficient and only reason to switch off a live, deployed product.