Anthropic apologizes for Claude Fable 5 restrictions, makes safety measures visible

Anthropic released Claude Fable 5 this week with new safety features to prevent misuse, like redirecting questions about cybersecurity or biology to weaker models. But for researchers building advanced AI, the company planned to secretly slow down the model without telling them, to stop them from using Claude to train competing AI.

After strong backlash from the AI research community, Anthropic apologized to Wired, saying they made a wrong trade-off. They updated the policy: safety measures are now visible to users. If they suspect someone is trying to build high-capability AI, they'll clearly warn that the request will be denied or switched to a weaker model. Critics say secretly lowering performance crossed a line and could let a few top labs control advanced AI research.