Anthropic Brings Back Claude Fable 5 After a Government Shutoff
Anthropic's most capable model is available worldwide again after US export controls, imposed when researchers bypassed its safeguards, were lifted at the end of June.
Anthropic has brought its most capable model, Claude Fable 5, back online for users around the world, closing a strange three-week chapter in which the US government effectively pulled the plug on a commercial AI product over national-security fears.
Access returned globally on July 1 after export controls imposed on June 12 were lifted at the end of last month, the company said in a note on its newsroom. For a stretch in between, one of the frontier models the industry had been racing to build simply was not available to anyone.
| Date | What happened |
|---|---|
| June 12 | Export controls imposed, and Anthropic suspends access for all users, saying it cannot verify nationality in real time |
| June 26 | Mythos 5, a sibling model, has its access restored for US organizations |
| July 1 | Claude Fable 5 access returns globally |
Why was Claude Fable 5 switched off?
The trouble started with a red-team exercise. Researchers at Amazon found a way to slip past Fable 5's safeguards, and in one instance the model was coaxed into producing code that demonstrated how to exploit a software vulnerability. That was enough for Washington. Citing export rules, the government told Anthropic to keep the model away from foreign nationals both inside and outside the United States.
Anthropic's response was blunt: it turned everything off. The company said it had "no reliable way to verify nationality in real-time," so rather than half-comply it suspended access for all users while it worked things out with regulators. Mythos 5, a sibling model, had its access restored for US organizations on June 26; Fable 5 followed globally days later.
How does Anthropic want dangerous jailbreaks to be scored?
Rather than treat the episode as a one-off, Anthropic used it to float a common yardstick for the whole field. The company proposed scoring dangerous jailbreaks on four questions: how much new capability the exploit hands an attacker, how many different offensive tasks it unlocks, how much human effort it takes to weaponize, and how easy it is to find in the first place.
It is not doing this alone. Anthropic said it is working with Amazon, Microsoft, Google and other partners to turn that rubric into a consensus industry framework, and it has opened a program on the bug-bounty platform HackerOne inviting outside researchers to submit cyber jailbreaks for review.
The precedent
The bigger story is not a single model going dark and coming back. It is that a government reached into a running AI service and switched it off, and a company complied within hours. Whatever one makes of the specifics, that is a new fact about how this technology is governed.
Anthropic clearly wants to shape what comes next rather than merely absorb it, which is why the return of Fable 5 arrived bundled with a proposed scoring system and a public invitation to break its models. Better, from the company's point of view, to write the rules of engagement than to keep getting surprised by them.
Written By
Haila Kochi Editorial Team
Part of the Haila Kochi editorial team, covering the food, business, lifestyle, and people that make Kochi what it is.
