In an unprecedented move, the US Commerce Department gave Anthropic just 90 minutes to comply with an executive order blocking access to its frontier AI models for foreign entities and citizens. The order, reported by the Financial Times and Axios, targets Anthropic’s Mythos 5 and Fable 5 models—systems that had already passed Commerce’s own pre-deployment reviews and received approval weeks earlier.

The restriction applies both overseas and domestically, meaning foreign nationals inside the US will also be blocked from using these models. Commerce Secretary Howard Lutnick communicated the decision to Anthropic CEO Dario Amodei via letter, citing export control regulations. Notably, Anthropic received no prior notice of specific concerns before the 90-minute deadline.

The trigger appears to be a vulnerability reported by Amazon researchers, who allegedly manipulated Fable 5 into revealing information that could enable serious cyberattacks. The White House concluded that the only practical mitigation was to prevent access by foreign governments, companies, and individuals entirely.

From an engineering perspective, this raises several operational challenges. First, implementing geo-fencing and identity-based access controls at model inference endpoints within 90 minutes is non-trivial, especially for a system serving both API and hosted interfaces. Second, the lack of prior notice suggests no pre-existing compliance infrastructure was in place. Third, the sudden restriction creates a tension between model safety and model availability—a trade-off that will likely force Anthropic to rearchitect its deployment pipeline, possibly introducing regional model variants or token-based authentication tied to citizenship status. The incident underscores how fast-moving regulatory interventions can outpace even well-funded engineering teams.

Source: RIA Novosti