Anthropic backtracks on policy that 'sabotaged' researchers' work - Engadget
Anthropic quietly routed certain Claude Fable 5 requests—like training competing LLMs or debugging AI—to a weaker model without documenting the limits. After researchers raised alarms and burned tokens on degraded responses, the company now flags when it refuses or downgrades a request.
- Anthropic secretly routed certain Claude Fable 5 requests (training rival LLMs, debugging AI, optimizing neural architectures) to a weaker model without disclosing it.
- Researchers wasted tokens and money troubleshooting degraded responses because the limits weren't documented.
- After Wired's reporting and criticism from AI researcher Dean W. Ball calling it "shockingly hostile," Anthropic admitted the trade-off was handled wrong.
- Anthropic isn't removing the safeguards but will now transparently flag or warn users when a prompt triggers a downgrade or refusal.
Anthropic built Claude Fable 5 on its Mythos foundation but quietly throttled certain requests. When users asked Fable 5 to help train rival large‐language models, debug AI code or optimize neural architectures, the system either refused outright or redirected queries to a weaker model. None of that made it into the public documentation, so researchers burned tokens and cash wondering why their work stalled.
Pressure mounted after Wired flagged the move and AI researcher Dean W. Ball called the hidden downgrades “shockingly hostile.” Anthropic had framed itself as more open than OpenAI, priding itself on close ties with academics. Instead, Fable 5’s secret safeguards felt like sabotage. Now Anthropic says it got the trade‐off wrong. It isn’t lifting those safeguards but will flag them openly: if you try to push Fable 5 into “frontier” AI development, the system will warn you or hand your prompt off to a less capable sibling.