Anthropic's Claude Fable 5 secretly downgraded AI researchers without telling them

Anthropic's launch of Claude Fable 5, a muzzled version of its top-tier Mythos model (part of Project Glasswing), has ignited controversy over hidden safety safeguards. While Anthropic clearly restricted bioweapon and chemistry requests by downgrading to Opus and notifying users, it applied a silent downgrade for researchers working on frontier AI models and advanced chip designs. Users were not told when the switch occurred, leading many to believe they were testing Fable's full capabilities when in fact they received Opus-level responses. The downgrade was mentioned only in Fable and Mythos System Card (319 pages), invisible in the user interface.

The backlash was swift: Fortune called it "secret sabotage," and Wired reported it could hinder AI research. Cybersecurity experts like Rob T. Lee (SANS Institute) warned that the same guardrails that block malicious actors also prevent legitimate defensive research. He experienced the silent downgrade while trying to build a digital forensics tool. Critics argue that the lack of transparency undermines trust, even if the restrictions were well-intentioned. The incident highlights the tension between safety and open research in frontier AI.