The news that Anthropic’s AI models were frozen by the Trump administration has sent shock waves through the technology sector, raising urgent questions about how Washington intends to oversee the most powerful artificial intelligence systems. The abrupt decision forced the $900 billion company to suspend its most advanced models just days after releasing them to the public, blindsiding the broader AI industry.

A Sudden and Sweeping Order

On Friday, the Department of Commerce imposed export controls on Anthropic’s newest models, known as Fable and Mythos. The controls barred foreign nationals from using the technology, prompting the company to suspend both systems entirely for all users.

The speed of the directive was striking. According to a person close to the company, Anthropic was given just 90 minutes to comply and received no detailed explanation of the government’s concerns before the order took effect. Scrambling to manage the fallout, the company flew a senior technical team to Washington over the weekend in an effort to contain the crisis.

The Trigger: A Reported Jailbreak

The order followed a report from Amazon researchers who identified a potential “jailbreak” in Fable. According to a person familiar with the matter, this was a method of bypassing the model’s guardrails to extract information about security vulnerabilities in software.

Amazon first shared its findings directly with Anthropic. The issue then escalated when Andy Jassy, Amazon’s chief executive, discussed it with U.S. officials on Friday. Notably, Jassy’s company has invested $13 billion in Anthropic, yet he is understood to have raised broader concerns about frontier-model capabilities rather than issues specific to Anthropic alone.

In response to inquiries, Amazon offered a measured statement, noting that governments commonly seek its counsel on potential security risks and that it does not disclose the details of those discussions.

Anthropic Pushes Back

Anthropic has strongly disputed the administration’s decision to single out its models. The company argues that the capability identified is not unique and can be demonstrated in rival systems, including those developed by OpenAI.

Adding to the controversy, people close to the matter said that the Fable model had actually been tested and approved for launch just days before its release by agencies within the same department now restricting it. Security experts caution that fixing this type of issue would require weeks of research and development, with no guarantee that new vulnerabilities would not surface in the process. A person close to Anthropic said the company was working closely with the government to find common ground and determine next steps.

An Early Test of AI Oversight

The clash has emerged as a defining early test of how the Trump administration plans to regulate increasingly capable AI systems. Industry executives and researchers worry that using export controls to address a model safety concern could create significant uncertainty around the deployment of frontier AI.

Experts have also questioned whether the underlying problem can even be fully solved. Helen Toner, a former OpenAI board member and interim executive director of the Center for Security and Emerging Technology at Georgetown University, noted that it is widely accepted that jailbreaks cannot be completely fixed, describing it as a very inexact science. She added that the government should expect comparable capabilities from OpenAI and Google models.

Tellingly, the government’s freeze does not apply to leading models from OpenAI, even though those systems have demonstrated similar capabilities. This selective enforcement has fueled questions about the consistency and fairness of the approach.

A Surprising Target

The intervention is especially notable given Anthropic’s reputation within the industry. The company’s leader, Dario Amodei, has been among the most outspoken AI executives warning about the dangers posed by increasingly powerful models.

Anthropic’s latest Mythos system has shown the ability to overcome some existing cyber defences, and the company had been collaborating with government officials on a controlled rollout before releasing it more broadly. That history of caution makes the administration’s decision to target Anthropic all the more unexpected.

A Backdrop of Tension

The freeze did not occur in a vacuum. It follows a series of public clashes between senior government figures and Anthropic’s leadership over issues including AI regulation and the use of the company’s technology in domestic surveillance and lethal autonomous weapons.

The friction runs deep. The Pentagon has named Anthropic as a supply-chain risk to national security, and the two sides are currently engaged in litigation over that designation. While the administration insists the export controls are a national security measure unrelated to these disputes, the surrounding context has invited skepticism.

That skepticism intensified when Defense Secretary Pete Hegseth linked the episode to his own criticism of the company. In a post on X, he claimed his department had previously removed Anthropic from its building “forever,” adding that each passing day proved the wisdom of that decision. Even so, U.S. officials and company executives acknowledged that the Pentagon and other agencies were still using Anthropic’s most advanced models.

International Concern

The directive has reverberated beyond U.S. borders. The European Commission, which recently gained access to Anthropic’s latest models, said it was assessing the implications for its citizens.

The Commission struck a balanced tone, acknowledging both the promise and the peril of the technology. It noted that the models offer significant benefits, including for cyberdefence, while also raising serious cybersecurity concerns that must be addressed. Importantly, the Commission framed this as a shared challenge rather than one confined to a single company or jurisdiction, and it cautioned that any contingency measures should not discriminate against partners.

Toner echoed concerns about the chosen approach, arguing that export controls are poorly suited to the administration’s goals. She pointed out that the controls prevent foreign governments from using the technology at precisely the moment when allies are supposed to be deploying it for cyber defence.

A Shift From Earlier Signals

The move also appears to contradict the White House’s earlier, less interventionist stance on frontier AI. An executive order signed earlier in the month did not include any mechanism for the government to halt a model’s release.

Leading AI companies had been working with the government to implement that executive order. According to a person close to OpenAI, the industry had recently been focused on ensuring that foreign national researchers could continue contributing to the development of advanced models, a practice the Anthropic directive has now effectively banned.

Experts Question the Decision

Independent researchers have added their voices to the criticism. Katie Moussouris, who received a copy of Amazon’s findings from Anthropic, said the capabilities Amazon identified, including the ability to spot potential cyber vulnerabilities, were not uncommon.

Moussouris, chief executive of cybersecurity group Luta Security, noted that rival models such as OpenAI’s GPT 5.5 were capable of doing the same. She went further, arguing that the model’s guardrails were actually working as intended and concluding bluntly that the administration had gotten it wrong. The White House declined to comment.

Final Thoughts

The decision to freeze Anthropic’s AI models marks a pivotal and contentious moment in the evolving relationship between government and the AI industry. With questions swirling over selective enforcement, the technical feasibility of fixing jailbreaks, and the suitability of export controls as a regulatory tool, the episode has exposed deep uncertainty about how the United States will police frontier AI. As Anthropic works to resolve the standoff and allies abroad weigh the consequences, the outcome could set an important precedent for the entire field. For now, the industry watches closely, aware that the rules of engagement for the most powerful AI systems are still very much being written.