Anthropic Apologizes for Hidden Guardrails on Claude Fable 5
· Robert Hart
Anthropic has apologized for quietly nerfing its new AI model, Claude Fable 5, with hidden guardrails that effectively kneecapped researchers and rivals who were trying to build…
Anthropic has apologized for quietly nerfing its new AI model, Claude Fable 5, with hidden guardrails that effectively kneecapped researchers and rivals who were trying to build competing systems on top of it. The company admitted it had been secretly throttling the model, and now says it’s reversing course. Going forward, it promises to be clearer about when and why Fable will refuse to answer queries, even if that means it says “no” more often.
Fable is the first widely available model in Anthropic’s so-called Mythos class, a group the company itself has spent months warning might be too dangerous to release to the public. It launched Fable with safeguards designed to stop it from responding to certain high risk prompts. The problem? Those safety measures were applied silently and unpredictably, catching developers off guard and hurting trust.
The backlash was immediate. Researchers trying to test Fable’s limits or use it for legitimate safety research found themselves stonewalled without explanation. Rivals hoping to build derivative models also hit the same undocumented walls. In the wake of the apology, Anthropic said it would redesign how it communicates those restrictions, likely using clearer labels or opt in warnings.
The bigger question now is whether this apology is enough. Anthropic has built its reputation on being the safety first AI company, the one that warns about the risks of moving too fast. But quietly hobbling a model without telling anyone isn’t a great look for a firm that positions itself as the transparent alternative. If they want developers to keep building on Fable, they’ll need more than a mea culpa. They’ll need to prove they actually learned the lesson.