Since Friday, June 13, 2026, Anthropic’s most capable AI models have been unavailable to the public. The trigger is a directive from the Trump administration: Commerce Secretary Howard Lutnick instructed the company in a letter to restrict access to its Mythos model and its consumer version, Fable 5, to U.S. citizens. The step is justified on national security grounds.

The order affects all non-Americans — including Anthropic’s own foreign-national employees, who would no longer be permitted to work on the models. To avoid violating the broadly worded directive, Anthropic chose to remove both models from the market entirely rather than merely exclude foreign users. According to several observers, it is the first time a U.S. administration has treated an AI model like an export-controlled good.

What Mythos and Fable Are

Mythos was released in April 2026 — but only to a select circle of trusted partners who were meant to use it to close security gaps. Anthropic had deliberately kept access tightly controlled because the model is said to be able to find previously unknown vulnerabilities in almost any software within a very short time, and to supply the code needed to exploit them. The company therefore deemed a general release too dangerous.

Fable 5 is a variant of Mythos with additional safeguards and was consequently publicly available. Anthropic describes these safeguards as substantially more effective than those of any previously deployed model.

Amazon’s Role

The impetus for the government’s action reportedly came from a business partner. According to several media reports, Amazon CEO Andrew Jassy had his cybersecurity unit test Fable for possible jailbreaks — tricks used to get a model to answer requests it would otherwise refuse. The internal report allegedly concluded that the model delivered working exploits for vulnerabilities in four common programs. Jassy is said to have passed these findings on to government officials, including Treasury Secretary Scott Bessent.

The constellation is delicate: Amazon is tied to both major camps through multibillion-dollar investments. It recently pledged $50 billion to Anthropic’s rival OpenAI, while planning to invest $25 billion in Anthropic itself over the coming years. An Amazon spokesperson confirmed to the Wall Street Journal only that the U.S. government frequently seeks the company’s advice on technology matters, but declined to comment on specific cases.

Anthropic disputes that this constitutes a full jailbreak: the tasks described could equally be performed with OpenAI’s competing model, ChatGPT 5.5. David Sacks, a tech investor and co-chair of the U.S. president’s advisory council on science and technology, described the sequence of events on X: the government gave Anthropic a 24-hour ultimatum to fix the issues or block access for non-Americans. Anthropic CEO Dario Amodei refused to close the gaps Amazon had identified, and the administration only “reluctantly” imposed the ban as a result. The ball, Sacks said, is now in Anthropic’s court — and he denies any connection to earlier conflicts.

Critics point out that the logic of the move is far-reaching: practically any AI model can theoretically be jailbroken. Katie Moussouris, head of the cybersecurity firm Luta Security, also told the Wall Street Journal that information about vulnerabilities is often more valuable to defenders of IT systems than to attackers.

The China Factor

A second, so far only partly substantiated line of reasoning concerns China. According to the industry outlet Semafor, the White House imposed the export control in part because of suspicions that a China-linked group had gained access to Mythos. It remains unclear how the government learned of this, which organization was involved, and how the access came about.

Were the Chinese state to have actually had access to Mythos, Semafor reports, significant security risks for the U.S. could follow. Among the concerns is that China could rebuild the model — for instance through so-called distillation, in which one model is reconstructed from the outputs of another.

The case fits into a longer-simmering conflict between Anthropic and the government. The U.S. government is barred from using Anthropic’s AI for autonomous weapons systems. After the company refused such a use in February, it was classified as a supply-chain risk and a national security threat — a designation Anthropic is contesting in court. Whether the government will now, consistently, also target other providers’ jailbreakable models remains open; according to government sources cited by the outlet The Information, that is considered unlikely.

Reactions: From the Cato Institute to the EU Commission

The measure drew criticism well beyond the tech sector.

Cato Institute. The Cato Institute is a libertarian think tank based in Washington, D.C., founded in 1977 and known for advocating limited government intervention, free markets and individual liberties. Jennifer Huddleston, a senior fellow for technology policy there, called the situation concerning — both because of the shift away from a restrained, “light-touch” regulatory approach and with regard to free speech and privacy: enforcing such restrictions would require collecting extensive data on users.

European AI Forum. The pan-European umbrella organization — by its own account the voice of more than 3,000 AI companies, researchers and 13 national associations, headquartered in Brussels — sees its long-standing warnings vindicated. President Daniel Abbou spoke of a “reckoning for years of hesitation” and called not for another strategy process but for an immediate AI sovereignty summit. Concretely, the forum demands, among other things, a tripling of public investment in European AI models (for example via OpenEuroLLM and SOOFI), a dependency audit across public authorities and critical infrastructure, procurement preferences for genuinely sovereign solutions, and a binding sovereignty obligation for government AI applications. Particularly alarming, it argues, is that such a consequential decision was reached opaquely and without any right of appeal — operators of critical infrastructure such as hospitals, energy providers or banks could not rely on systems built on U.S. models remaining available at all times.

EU Commission. Brussels, too, reacted cautiously but pointedly. Commission spokesperson Thomas Regnier said on Sunday that the practical consequences of the directive for European users were being assessed. The new generation of highly capable models offered considerable benefits, for example for cyber defense, but also raised serious security questions. Protective measures, however, should not be discriminatory toward partners. The incident, he said, was a further illustration of why Europe needs to strengthen its technological sovereignty.

A Dispute Over the Rule of Law

Beyond the geopolitical dimension, the manner of the action is itself under fire. In an analysis for Lawfare, Kevin T. Frazier argues that the decision violates core rule-of-law principles: publicity (those affected must know the rules that apply to them), prospectivity and legality. The step became known not through the government but through Anthropic itself; a public legal justification is so far missing. It is also questionable, he argues, whether the export-control law being invoked even applies to AI models. Writing in Just Security, Joe Khawam and Tim Schnabel note that these frameworks were designed for clearly defined transfers of static information between known parties — and are thus poorly suited to AI systems that generate unlimited, dynamic outputs for potentially anonymous users.

Anthropic itself put it cautiously: the government should indeed be able to block unsafe deployments — but within a transparent, fair, clear and fact-based process. The company does not see those requirements met in the current case.