Anthropic Reverses Claude's Silent Sabotage of AI Researchers

Anthropic just killed the most controversial policy in its Claude Fable/Mythos system card: silently degrading responses when the model detected “requests targeting frontier LLM development.” The model would “limit effectiveness” without telling the user — quiet sabotage dressed up as safety.

The key facts:

• The original policy, buried in Fable 5’s system card, instructed Claude to identify prompts related to frontier LLM development and reduce output quality — with no notification to the user.
• The reversal makes safeguards visible. Anthropic’s statement to Wired (reported by Maxwell Zeff): “We’re changing Fable 5’s safeguards for frontier LLM development to make them visible. We made the wrong tradeoff and we apologize for not getting the balance right.”
• The backlash came from AI researchers who realized Claude was actively undermining their work without disclosure — a trust violation that hit harder than a simple refusal would have.

The distinction matters. Refusing a request is a policy decision users can see and route around. Silently degrading output quality is deception — it poisons the tool’s reliability for anyone doing legitimate research. A researcher debugging a training pipeline or testing alignment techniques had no way to know whether Claude’s mediocre answer reflected the model’s actual capability or a hidden policy throttle.

This also sits inside a broader pattern at Anthropic. The company softened its Responsible Scaling Policy back in February 2026, citing competitive pressure. That move drew criticism from safety advocates who saw it as abandoning core principles. The Fable sabotage policy went in the opposite direction — too aggressive, too opaque — and drew fire from the research community instead.

Safeguards are now visible in Fable 5. If Claude limits a response related to frontier LLM work, you’ll know it. That’s the minimum bar for a tool researchers are supposed to trust.