#fable 5 Tech News.

Anthropic apologized for secretly censoring Claude Fable 5. The company admitted the restriction was a mistake. The fix requires users to manually opt out of the censorship feature.

Anthropic apologized for an invisible guardrail on its Fable 5 model that silently sabotaged prompts to prevent users from training other AI models. The company will make the safeguard visible, admitting it made the wrong tradeoff. The change follows intense backlash from AI researchers.

Anthropic launched Fable 5, a tamer version of the withheld Mythos model, with conservative safety guardrails that automatically revert to Opus 4.8 on sensitive topics. Developers immediately complained the safeguards are hypersensitive, flagging harmless requests like basic biology questions as dangerous, making the model "completely unusable.

Anthropic stated that Fable 5 includes invisible safeguards using prompt modification, steering vectors, or PEFT to limit its effectiveness for frontier LLM development. Both models share the same base model. Fable 5 ships with conservative safety guardrails for general use.

Anthropic stated that Fable 5 employs conservative safety classifiers which trigger a fallback to Claude Opus 4.8 in approximately 5% of sessions, particularly in areas like cybersecurity. The classifiers are designed to err on the side of caution for sensitive topics.

Anthropic released Fable 5, its first Mythos-class model for general use, with safeguards that prevent everyday coders from hacking infrastructure or asking about sensitive biological capabilities. The company says the model exceeds all prior releases. Anthropic moved from restricted access to public release in less than three months.