Amazon’s Security Warning Led to US Ban on Anthropic AI Models

A Wall Street Journal report says Amazon CEO Andy Jassy went directly to the White House after Amazon security researchers found a way to bypass safeguards in Anthropic’s models. That warning helped trigger an extraordinary US government order that blocked access outside the United States to two of Anthropic’s AI models, including its most advanced systems.

According to the report, the issue began when Amazon researchers identified a possible way to extract information from the model Fable 5 that could be used for cyberattacks. Instead of going through the usual vulnerability disclosure process with Anthropic, Jassy contacted the administration himself. Soon afterward, Commerce Secretary Howard Lutnick sent Anthropic CEO Dario Amodei a formal letter ordering the company to cut off access outside the US.

The technique discovered in Anthropic’s models is known as a jailbreak, meaning an attempt to defeat an AI system’s guardrails so it will do prohibited tasks. Anthropic said this was not unique to its models and also exists in rivals such as OpenAI’s GPT-5.5. Even so, Anthropic had to suspend access broadly, and some employees outside the US also lost access to tools they had helped develop.

The episode has sparked debate in the US because Amazon is both a major Anthropic investor, with billions of dollars committed, and a direct competitor in AI. Amazon also supplies cloud infrastructure and chips for training Anthropic’s models. The move has raised suspicion that security concerns may have masked a competitive effort. The case also comes amid tensions between Anthropic and the Trump administration, after federal agencies restricted use of the company’s products in February. The Commerce Department has not yet detailed the specific threat, but the decision is already being seen as a turning point in AI regulation.