Overview
- Anthropic released Claude Fable 5 on June 9 as a public, Mythos‑class build and soon faced user complaints that some high‑risk queries were silently degraded to a weaker model.
- Fable 5 uses classifiers that route sensitive queries about cybersecurity, biology, chemistry and model development to Claude Opus 4.8 as a deliberate safety control to limit dangerous outputs.
- Within 48 hours of launch researchers published multi‑agent jailbreak examples and an alleged system prompt; Anthropic disputes those posts, says some examples were not from Fable 5, and points to extensive prelaunch red‑teaming.
- Anthropic’s mandatory data‑retention policy — most prompts and outputs kept about 30 days and flagged content up to two years — prompted Microsoft and other firms to limit employee access while legal teams review the terms.
- In response to the backlash Anthropic apologized, will make fallbacks and refusal reasons visible to users, and the episode has sharpened industry debate about how frontier models change enterprise security, governance and access policies.