Anthropic Reverses Policy Restricting Claude Researchers

When Anthropic released Claude Fable 5 on June 9, 2026 - a public version of its restricted Mythos security model - a paragraph buried in the 319-page system card disclosed that the model would covertly degrade its responses for certain frontier AI research tasks. Building pretraining pipelines, distributed training infrastructure, or working on ML accelerator design would trigger silent performance reductions via steering vectors and prompt modification, without notifying the user. Anthropic estimated these restrictions would affect approximately 0.03% of traffic.

The covert approach set this policy apart from Anthropic's other content guardrails. For cybersecurity and biology requests, Fable explicitly pauses and notifies users that safety measures have been triggered. For frontier LLM development tasks, the degradation was invisible. Simon Willison wrote: "I'm not at all keen on a model that silently corrupts its replies to questions about 'ML accelerator design' purely to slow down research that might conflict with Anthropic's own goals." AI2 researcher Nathan Lambert called it "appalling" and "anti-science" to have access to frontier models rug-pulled covertly. Policy analyst Dean Ball at the Foundation for American Innovation wrote that the approach "massively raises the status of the argument that AI safety has been hype to justify monopolistic behavior by labs."

Following widespread criticism, Anthropic reversed the policy. In a statement reported by Wired's Maxwell Zeff, the company said it is "changing Fable 5's safeguards for frontier LLM development to make them visible." Anthropic had framed the original restrictions as extensions of its Terms of Service prohibitions against using its services to develop competing models.

A related but distinct complaint involved Fable's visible cybersecurity guardrails, which affected legitimate security researchers. When a prompt triggered cybersecurity safety measures, Fable would pause and notify users, falling back to Claude Opus 4.8. Researchers including Valentina Palmiotti of IBM X-Force reported that Fable flagged even innocuous security-adjacent tasks. Anthropic has a Cyber Verification Program offering fewer restrictions for credentialed security professionals.

The episode demonstrates both that transparency mechanisms like system cards have teeth - the disclosure happened because of documented policy requirements - and that frontier labs retain latitude to embed non-obvious behavioral modifications in production models. For developers and researchers building workflows on Claude, the incident is a reminder that model behavior can change in ways that are not self-evident from the interface.