Trump Signs AI Order: Powerful Models Under Government Scrutiny Before Release
In recent hours, Donald Trump signed an executive order requiring artificial intelligence companies to make their most powerful models available to the federal government up to 30 days before public release, for an assessment of their offensive capabilities in cyberspace. The signing occurred privately, without the presence of Silicon Valley CEOs initially expected.
Trump was supposed to sign a stricter version of the order on May 21, but the ceremony was postponed due to pressure from the industry, including from investor David Sacks, who previously led AI and cryptocurrency policies at the White House. The original draft allowed the government up to 90 days to review a model, while some industry operators reportedly pushed for a window of just two weeks. The signed text limits this to 30 days.
What the Order Entails
The measure tasks a group of federal agencies with developing a classified evaluation process to measure the advanced cyber capabilities of a model and establish the threshold beyond which it is designated as a "covered frontier model". The designation is the responsibility of the NSA director. Based on this, the government will create a voluntary framework allowing developers to grant early access to designated models, for no more than 30 days, to a select group of “trusted partners” chosen along with the authorities.
The order explicitly states that nothing in the text authorizes the creation of a licensing obligation, pre-approval, or government permission for the development or release of models. Cooperation remains voluntary. Simultaneously, the measure tasks the Treasury with creating a coordination structure on software vulnerabilities and instructs the Justice Department to treat crimes committed using AI for unauthorized access to computer systems as a priority.
The Shadow of Mythos
The elephant in the room is Claude Mythos, the model from Anthropic that has identified thousands of zero-day vulnerabilities across all major operating systems and browsers. We have previously discussed it in relation to Project Glasswing, the initiative where Anthropic chose not to release the model to the public and instead entrusted it to a consortium of companies for defensive purposes only. It is exactly these types of capabilities that the order aims to intercept before a model hits the market, with an eye toward protecting critical infrastructure: rural hospitals, local banks, and utility companies.
However, the relationship between the administration and Anthropic remains tense. The Pentagon has classified the company as a supply chain risk after its refusal to grant the military unlimited access to its models, and Anthropic has sued to overturn this designation; the lawsuit is ongoing.
Nonetheless, the order represents a departure from the stance held by the administration so far, which had imposed very few constraints in last summer's AI action plan: for the first time, the White House acknowledges that frontier models pose concrete security risks. However, the Center for Democracy and Technology has described the mechanism as "opaque" because it offers the public little visibility into the evaluation process, and a lack of transparency leaves room for discretionary use of release power.
The text lacks operational details: it sets deadlines for writing directives but leaves the construction of the voluntary framework to the agencies. How much a mechanism based on spontaneous adherence and lacking obligations can truly bind the most advanced laboratories is a knot that the measure, in its current state, is not yet able to untie.