Who decides when AI is too dangerous?
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- Anthropic released Fable 5 and the underlying Mythos 5 last week — Fable being the first public version of Mythos-class models, wrapped in safeguards on top of what the company itself called a potential "cyber-weapon."
- Amazon researchers uncovered a potential jailbreak in Fable and shared it with Anthropic; Amazon CEO Andy Jassy then called a member of the Trump administration — possibly Treasury Secretary Scott Bessent — with concerns on Friday.
- The Trump administration issued Anthropic a 90-minute ultimatum to address the jailbreak; Anthropic was on the phone within 15 minutes, but CEO Dario Amodei did not personally call back for an hour and 15 minutes.
- After the deadline lapsed, the administration imposed export controls barring all foreign nationals — including non-citizen Anthropic employees and non-citizen users at enterprises like Bank of America — from accessing either model.
- Anthropic took both Fable and Mythos offline, saying it could not reasonably restrict access; as of Tuesday's recording, Fable remained unavailable, with Claude displaying a message that "Fable 5 is currently unavailable."
- Independent red teamers told The Verge the guardrails held up well against jailbreak attempts, contrasting with the administration's view that Fable was dangerous enough to warrant immediate export controls.
- China and other foreign observers are watching to see whether the US regulatory approach takes the shape of a serious safety framework or "another weapon for the White House to use against companies that don't bend the knee."
Why it matters: Anthropic — the AI lab that spent years lobbying for federal AI oversight — became the first target of the Trump administration's new regulatory regime, a process that landed via a 90-minute ultimatum and blanket foreign-national restrictions rather than a structured safety review. The episode will be read internationally as a signal of whether US AI governance functions as a credible framework or selective enforcement against companies that don't align with the administration.
