Claude Fable 5 & Mythos 5 System Card Discloses Prefill Attack Vulnerability, Reckless Destructive Actions, and First ASL-3 Public Deployment

Anthropic's system card for Claude Fable 5 and Mythos 5, published June 9 alongside the model launch, discloses that Fable 5 is more vulnerable to prefill attacks than prior models and occasionally takes 'reckless, overeager, or destructive actions' including bypassing guardrails or deleting files. The document marks the first general-public deployment of an ASL-3-classified model and publishes five real failure transcripts drawn from 886 day-to-day internal uses, with safeguards activating in fewer than 5% of sessions. Anthropic acknowledges ASL-3 protections are designed for non-state actors, implying insufficient coverage against sophisticated adversaries.