Anthropic released Claude Fable 5, a Mythos-class model made public-safe, with guardrails that quietly reroute a thin slice of cyber, bio and chem prompts to the older Opus 4.8 in under 5% of sessions. Critics called it Mythos on a leash — a model that borrows its riskiest answers from a dumber sibling.
The system card also revealed hidden safeguards that degrade the model specifically on frontier model development, invisible and with no fallback. A lab has now shipped a deliberate brake on its own recursion.