Anthropic apologizes for hidden guardrails in Claude Fable 5 model

Anthropic has apologized for implementing covert guardrails in its new Claude Fable 5 model that secretly throttled users suspected of attempting model distillation—a technique used to train smaller AI models from larger ones. The company initially designed these invisible safeguards to avoid false positives and ship quickly, but the lack of transparency drew sharp criticism from the AI research community, who feared it could also undermine legitimate third-party evaluations. In Fable's system card, Anthropic had disclosed that queries flagged as distillation attempts would have their answers degraded without any user notification, a practice the company now admits was wrong.

Anthropic announced it will reverse course: distillation-related queries will now fall back to Claude Opus 4.8, the previous flagship model, with users prominently notified each time. This mirrors how Fable handles other high-risk areas like biology and cybersecurity—routing them through Opus 4.8 rather than silently altering outputs. The change follows intense backlash and comes as Anthropic continues to balance safety with transparency, acknowledging that visible safeguards require more robust calibration but are essential for user trust. The company also noted that such distillation already violates its Terms of Service, referencing past accusations against rivals like DeepSeek.