The United States has moved to restrict Anthropic's Claude Fable AI model, and the stated reason cuts to the heart of a growing tension between government security demands and how AI companies operate.
According to David Sacks — cited by Yellow.com — Anthropic refused to patch a known jailbreak vulnerability in Claude Fable, and the US government responded by banning the model. A jailbreak, in plain terms, is a technique that tricks an AI into ignoring its own safety guardrails, potentially enabling harmful outputs. When a company declines to fix such a flaw, regulators face a choice: accept the risk or restrict access.
The fallout is already crossing the Atlantic. According to Reuters, the European Commission said it is assessing the practical implications of the US export control directive targeting Anthropic's models. Crucially, EU officials noted that such measures "should not be discriminatory against partners" — a signal that Brussels views the restriction as a potential trade irritant, not just a security matter.
Not everyone agrees the ban is sound policy. The R Street Institute, a free-market think tank, described the situation as "a bad idea applied badly," suggesting the approach itself — beyond any specific grievance with Anthropic — is poorly conceived.
The episode matters because it marks a rare moment where a government has publicly punished an AI lab not for what its model did, but for what the company refused to do — raising hard questions about who controls AI safety decisions and what happens when regulators and developers disagree.