U.S. move against Anthropic models sparks debate over AI as a security asset

The U.S. government has for the first time moved beyond restricting access to AI infrastructure and instead limited access to the models themselves, blocking foreign nationals from using Anthropic’s new Claude versions, Mythos 5 and Fable 5. The decision, which reportedly followed concerns that the models’ safety controls could be bypassed and that their ability to find security flaws might aid hostile actors, has jolted Silicon Valley and raised fears of a precedent that could reshape the industry.

According to Axios, administration officials tried to persuade Anthropic to delay the launch. When the company refused, unusual export restrictions were imposed that barred foreign citizens, including those physically in the United States, from accessing the models. To comply, Anthropic disabled the models for everyone. The dispute centers on a capability that is useful for defenders and attackers alike, finding vulnerabilities in code and computer systems.

Amazon, Anthropic’s biggest investor and one of its closest strategic partners, became a focal point in the controversy. Reports in The Wall Street Journal, Reuters and Axios say Amazon CEO Andy Jassy spoke with Trump administration officials and shared findings from Amazon researchers about Anthropic’s models. While there is no indication Amazon asked for a ban, its involvement in the discussions that preceded the review intensified the backlash.

Anthropic rejects the government’s assessment, saying the risk has been exaggerated and that no evidence was presented of a systematic ability to bypass safety protections. The company says it ran months of security testing with outside researchers and government officials, and none succeeded in breaking the models’ defenses. Critics and supporters alike see the episode as a test of whether advanced AI should be treated as sensitive national security infrastructure rather than a normal cloud service.

Analysts say the move could accelerate efforts in Europe and Asia to build independent models, and some warn it may slow U.S. innovation by giving rivals time to narrow the gap. Others think Anthropic will soon release a modified version and that the uproar is partly online noise. For Israel, the episode has renewed calls for national AI capabilities, including a local language model, because core systems may depend on foreign providers and U.S. policy.