Trump Backs Off Mandatory AI Model Review After Anthropic's Mythos Controversy

President Donald Trump pulled out at the last minute from signing an executive order that would have forced developers of the most powerful AI models to submit them to government review before public release. The push for the rules was triggered by Anthropic's Claude Mythos model, which the company itself deemed too dangerous to release freely.

A model deemed too dangerous

Anthropic said Claude Mythos Preview could detect thousands of critical software vulnerabilities, and that its offensive cybersecurity capabilities emerged as a side effect of broader advances in reasoning and code generation. The company decided not to release the model publicly and instead launched Project Glasswing, a program granting limited access to the tool to roughly 70 trusted organizations, which could use it to find and fix vulnerabilities in their own systems first.

One result of the program was Mozilla's discovery of 271 security bugs in Firefox, which the company managed to patch before information about them spread more widely. At the same time, Anthropic terminated a $200 million contract with the Pentagon after refusing to remove safeguards limiting the use of AI in autonomous weapons systems and mass surveillance. A federal judge called the Pentagon's conduct in the dispute "Orwellian," even though it later emerged that the National Security Agency (NSA) was using Mythos anyway to analyze vulnerabilities in Microsoft software.

A plan for mandatory review

The Mythos affair landed on the White House's desk. According to a New York Times report from May 5, 2026, the Trump administration was considering setting up a special working group of government officials and executives from Anthropic, Google and OpenAI to design a system of mandatory review for new AI models. Oversight was to fall to the National Security Agency, the Office of the National Cyber Director, and the Director of National Intelligence.

The proposal drew inspiration from the UK model, in which public institutions test advanced AI systems for safety and compliance with standards, and partly from the drug approval process used by the US Food and Drug Administration (FDA). It called for mandatory safety documentation, red-team testing for offensive capabilities, and a requirement to report models exceeding certain risk thresholds to a federal review body.

Signing called off

The signing ceremony for the executive order, scheduled for May 21, 2026, was canceled at the last minute. Trump told reporters in the Oval Office that he didn't like "certain aspects" of the document.

I don't want to do anything that would get in the way of that lead - Donald Trump, President of the United States, on the US's leading position in the AI race against China

The president argued that artificial intelligence "brings enormous benefits," and that mandatory model review could become a "blocker" for American companies. The decision effectively shelved a plan that would have required leading AI developers to submit their most powerful systems for government evaluation before public release.

A watered-down version and fights in Congress

Two weeks later, on June 2, 2026, Trump privately signed a much softer version of the rules, an executive order titled "Promoting Advanced Artificial Intelligence Innovation and Security." Instead of a mandate, it establishes a voluntary system under which model developers can submit new technologies for review by federal agencies, including the Treasury Department, the NSA, and the Cybersecurity and Infrastructure Security Agency (CISA), with a window of up to 30 days before a model reaches trusted partners on the market. The agencies are also to develop classified criteria defining which models qualify as "frontier AI models" based on cyber capabilities, with the framework due to be finalized by August 1, 2026. Senator Josh Hawley, a Republican from Missouri, has pushed for mandatory reviews to be established through legislation rather than an executive order alone.

Congress must now codify the White House's executive order into law - Brendan Steinhauser, Alliance for Secure AI

Critics point out that executive orders can be reversed at any time with another signature, whereas legislation would give the review system a more durable legal foundation. The administration, however, has not signaled support for such a step.

What it means for the industry and Poland

The episode shows how difficult it is for US policymakers to reconcile two goals: limiting the risks posed by increasingly powerful AI models and keeping pace with China in the race to develop them. The final, voluntary shape of the rules means companies like Anthropic, OpenAI and Google decide on their own whether and when to submit their models for government evaluation, a contrast that European and Polish observers note with the EU AI Act's mandatory requirements for high-risk models. For Polish companies using American AI models, this means the only hard safety backstop for now remains EU regulation, not US export controls or licensing oversight.