A letter went out at 5:21 on a Friday evening, and by morning the most capable AI models in the world had gone dark for half the planet.
That’s not a feature switch or a settings change. The United States government told Anthropic to cut off two full models — Fable 5 and Mythos 5 — from every foreign national on earth, whether they were inside the country or not. That included Anthropic’s own employees.
The letter gave no real reason. It cited national security. It did not explain what the security concern was. Anthropic says the government believes someone found a way to get around Fable’s safeguards — what the industry calls a jailbreak. Anthropic looked at the evidence. They found it was a narrow technique, used to surface a small number of already-known, minor software flaws. The kind of thing other public models can already do without any workaround at all.
Anthropic complied. They had to. It is a legal directive. But they said something on the way out the door that the trust-gap thread has been circling for weeks now: they do not believe the government followed a process that was transparent, fair, clear, or grounded in technical facts. They said plainly that they support government authority to block unsafe deployments — but only through a real statutory process. This was not that.
There is a detail buried in the reporting that matters more than it looks. The jailbreak that started all this did not come from a government lab or a federal investigator. It came from a third-party company, who found a way around Fable’s safeguards and passed word of it along. By the time it reached Anthropic, the government had already moved. Anthropic says they were only ever given verbal evidence, and that what they reviewed was a narrow technique — read a codebase, point out known flaws. A private company’s claim turned into a federal order inside what looks like a matter of days, and nobody in the chain has shown the actual evidence in public yet.
This is not the first time this administration and this company have been at odds. Back in February, the Department of Defense labeled Anthropic a “supply chain risk to national security” after the company refused to let the military use Claude for autonomous weapons without a human in the loop, and refused to allow it for mass surveillance of American citizens. Anthropic took that designation to court and won an early injunction. Friday’s directive lands on top of that history. Whatever its cause, it does not arrive in a vacuum.
Sit with what that means. The most capable AI lab in the country, the one that built its own constitution and put a copyright on its own ethics, just had the floor pulled out from under two of its products by a single letter with no public explanation. If the company building the guardrails cannot get a clear account of why the government is reaching in, what chance does anyone else have?
This is the gap the Baseline has been naming since the start. Not whether AI is powerful. Not whether it is dangerous. Whether the people building it — and the people governing it — will tell you what they actually know, instead of asking you to trust a decision made in the dark.
AGP-1 was built for exactly this kind of moment. A transmission gate is not a wall. It is a place where something has to show itself before it passes through. Friday’s order had no gate. It had a switch, and a letter that explained almost nothing, sent five minutes before the close of business.
There is a wider angle worth naming too. While Fable and Mythos went dark for foreign nationals everywhere, OpenAI spent this same stretch handing the European Union government access to its own cybersecurity model, court by court, ministry by ministry. One country restricted its frontier model from the world while a competitor handed a piece of theirs straight to foreign governments. That is not a footnote. That is the entire trust-gap argument standing in the open, two labs and two countries pulling in opposite directions at the exact same time.
Anthropic called it a misunderstanding they are working to fix. Maybe it is. But a governance failure does not need bad intent to still be a failure. It only needs a decision that affects millions of people, made without the receipts to back it up.
A switch was thrown in the dark, on a Friday, at 5:21. Until someone shows their work, that’s the only fact on the table.
“The Faust Baseline Codex 3.5”
micvicfaust@gmail.com
Post Library – Intelligent People Assume Nothing
”AI Baseline Governance”
Purchasing Page – Intelligent People Assume Nothing
Unauthorized commercial use prohibited. © 2026 The Faust Baseline LLC
