Anthropic scrambled after the Trump administration abruptly froze its $900bn AI models, Fable and Mythos, citing security risks. The Department of Commerce imposed export controls, barring foreign use and forcing the company to suspend both systems. This move, announced days after the models’ public release, has stunned the AI sector and ignited debates over regulatory overreach. A jailbreak flaw identified by Amazon researchers—allowing circumvention of safety guardrails—was cited as the trigger. Anthropic disputes the targeting, arguing similar vulnerabilities exist in rivals’ systems, including OpenAI’s models. The incident underscores tensions between industry innovation and government intervention in AI development.

The administration’s directive reflects broader unease about frontier AI capabilities. Amazon, which invested $13bn in Anthropic, initially shared findings about the jailbreak but kept details confidential. Andy Jassy, Amazon’s CEO, raised concerns about systemic risks rather than Anthropic-specific flaws. Meanwhile, Anthropic’s leadership, including CEO Dario Amodei—a vocal critic of AI risks—contends the freeze is disproportionate. Security experts warn that patching such flaws is technically challenging and temporary. The Pentagon’s classification of Anthropic as a national security risk further complicates the situation, despite ongoing use of its models by US agencies. The European Commission also flagged cybersecurity concerns, emphasizing the need for global cooperation rather than unilateral actions.

The clash highlights a pivotal moment for AI governance. While the White House had signaled a lighter touch earlier this month, this enforcement action sets a precedent for using export controls to address safety issues. Critics argue this approach stifles innovation and creates uncertainty for investors. For Anthropic, the fallout includes operational delays and strained relations with regulators. The company is now deploying a technical team to Washington to negotiate terms. Helen Toner, a former OpenAI board member, called the export control mechanism “poorly suited” for addressing AI security gaps. As the debate intensifies, the episode raises urgent questions about balancing technological progress with national security priorities. The immediate focus remains on resolving the jailbreak flaw, though experts caution new vulnerabilities could emerge.