US government forces Anthropic to pull Fable 5 and Mythos 5 — a landmark regulatory precedent for frontier models

On June 12, 2026, a US government directive forced Anthropic to disable Claude Fable 5 and Mythos 5 for all customers worldwide. The stated reason: a jailbreak risk enabling users to bypass the model's safety protections. Anthropic is complying but publicly contesting the proportionality of the measure — the identified vulnerabilities are described as minor and comparably present in OpenAI's GPT-5.5, which received no equivalent directive. That asymmetric treatment is the first concrete signal that US regulation can strike selectively, without a uniform technical threshold.

The stakes extend well beyond Anthropic. If a potential — not confirmed — jailbreak vector is sufficient to justify a global recall of a widely deployed model, every frontier lab operating under US jurisdiction faces the same exposure. Anthropic said it plainly: this precedent could block any future frontier deployment. The question is no longer purely technical (what jailbreak tolerance is acceptable?); it is institutional — who decides, on what basis, and with what transparency.

Against that backdrop, the publication on Hugging Face of 953 Fable 5 traces including CoT data — captured before the suspension by Glint-Research and TeichAI — adds another layer. The directive removes the model, not the data. Fine-tunes are already anticipated. Regulation-by-model-withdrawal runs directly into the reality of an open-data ecosystem where artifacts outlast administrative decisions.