Anthropic Releases Claude Fable 5 and Mythos 5 — Then Pulls Them Under US Government Pressure
Tags AI · Anthropic · Frontier Models · Cybersecurity
Anthropic publicly released two new frontier models: Claude Fable 5, a general-purpose model, and Claude Mythos 5, a cybersecurity-specialized model derived from Anthropic's internal 'Mythos-class' systems. Fable 5 was a public version of Anthropic's most capable model, shipping with guardrails that blocked cybersecurity queries. Within days, Amazon researchers demonstrated jailbreaks circumventing those guardrails, and the White House ordered the models withdrawn. The episode demonstrated that frontier AI capabilities — particularly in offensive cybersecurity — are advancing faster than the policy frameworks designed to contain them, and that model guardrails cannot reliably prevent misuse of the most capable systems.
Technical significance
The Fable 5 / Mythos 5 release and withdrawal is the most visible demonstration yet that the gap between AI capability and AI governance is widening. Mythos 5's cybersecurity specialization means Anthropic built a model that can discover and chain zero-day exploits — a dual-use capability that has no precedent in public AI releases. The fact that jailbreaks were demonstrated within days of release, and that security researchers say perfect jailbreak prevention is impossible, suggests the industry is entering a phase where the most capable models may only be deployable under restricted access regimes.