Amazon Chief Warned Trump Officials About Anthropic AI Security Risks, Leading to Global Shutdown

Amazon CEO Andy Jassy was among several tech executives who raised concerns this week with senior Trump administration officials about security risks in Anthropic’s advanced AI models, according to a Reuters source cited in U.S. reports. The Wall Street Journal said Amazon engineers found the flaw but went directly to the White House instead of informing Anthropic, a move that helped trigger the government action.

The administration ordered Anthropic to block access for all foreign nationals, in the U.S. and abroad, to two new models, Fable 5 and Mythos 5. Anthropic responded by shutting access to the models worldwide on Friday night, saying the restrictions were imposed under export-control measures. The company, based in San Francisco and quietly pursuing a U.S. IPO, said the government told it there was a way to bypass its safeguards and use the model to find cyber vulnerabilities. Anthropic countered that the bypass exposed only “minor” flaws that other public models can also detect.

In a blog post, Anthropic said the issue involved a jailbreak loophole. The company had already warned about Mythos’s hacking capabilities and had limited its wider release, while earlier this week it launched a public version of Fable that it said included cyber protections. White House adviser David Sacks said on Saturday the order was issued “reluctantly” after CEO Dario Amodei “refused” to “fix the jailbreak vulnerability or take down the model.” He said the administration hopes Anthropic will fix the safety problem, after which the export controls would be lifted and Fable could return to general release.

The episode highlights a complicated Amazon-Anthropic relationship. The companies compete in AI, but Amazon also provides cloud services for training and running Anthropic’s models and is one of its investors, with a potential windfall from Anthropic’s upcoming IPO. Amazon declined to say whether it had discussed Anthropic’s models with officials, saying governments often consult it on possible security risks and it does not disclose those conversations. The dispute also comes after Anthropic refused earlier this year to let the U.S. military use its models for internal surveillance or fully autonomous weapons, leading to a government supply-chain watchlist entry due later this year. Defense Secretary Pete Hegseth wrote that the Pentagon had removed Anthropic from its building three months ago, adding, “Every day that goes by proves we were right.”