Le gouvernement américain ordonne à Anthropic de tuer Fable 5 sur un jailbreak n'importe quel modèle peut faire

At 5:21pm ET on June 12, the US government ordered Anthropic to suspend access to Fable 5 and Mythos 5 for all foreign nationals — including Anthropic's own foreign national employees — citing national security authorities. The directive didn't specify the security concern, but Anthropic learned it's based on a reported method to bypass the model's safeguards.

The Jailbreak That Broke a Deployment

The government demonstrated a technique that identifies a small number of previously known, minor vulnerabilities. Anthropic reviewed the report and found those vulnerabilities are simple and easily found by other publicly-available models, including OpenAI's GPT-5.5. No universal jailbreak — one that broadly bypasses safeguards — was found. No harmful outcome was disclosed. The government's evidence amounts to asking the model to read a codebase and fix flaws, something defenders do every day.

Anthropic built Fable 5 with a defense-in-depth strategy explicitly acknowledging that perfect jailbreak resistance isn't possible. They required 30-day customer data retention to quickly detect and shut down attacks — a policy that hurt customer relationships but was designed for exactly this kind of scenario. Thousands of hours of red-teaming with the US government, UK AISI, and third parties showed Fable's safeguards are substantially more effective than any previously deployed model.

The Precedent That Stops All Deployments

Anthropic is complying, but they're blunt about the implications: if this standard — a narrow, non-universal jailbreak that any model can produce — is grounds for recalling a commercial model used by hundreds of millions of people, then every frontier model deployment is at risk. No provider currently ships perfect jailbreak resistance. The government skipped the transparent, technical process Anthropic has publicly advocated for.

Users lose access to Fable 5 and Mythos 5 immediately. Anthropic calls this a misunderstanding and is working to restore access quickly. Meanwhile, every AI lab now knows: one unverified narrow jailbreak, delivered after work hours, can kill a product — even when the same capability is already available from every other model on the market.

Source: US Government directive to suspend access to Fable 5 and Mythos 5
Domain: anthropic.com