President Trump signed a new AI executive order on June 2, asking frontier AI companies to let the federal government review their models up to 30 days before release. The framework is voluntary. No company is required to participate, and the order explicitly states it cannot be used to create mandatory licensing or pre-clearance requirements. Reaction from across the political and technical spectrum has ranged from cautious welcome to outright dismissal, with very little in between.
The mechanics are still being worked out. The Cybersecurity and Infrastructure Security Agency, the National Security Agency, and the Treasury Department have 60 days from the order's signing to define what counts as a "covered frontier model," the category that would trigger the optional review. The review period itself can run up to 30 days, during which the government would have access to the model to assess its cybersecurity capabilities. The order also establishes an AI cybersecurity clearinghouse to track and address AI-related software vulnerabilities.
OpenAI has already said it will participate. George Osborne, the company's head of countries and a former UK finance minister, told CNBC that signing up was "quite right," and that frontier labs should not wait to be asked before proactively working with governments on safety and security. That position is consistent with OpenAI's approach in Washington: engage cooperatively on governance frameworks that do not impose hard constraints, and help shape what those frameworks look like.
The motivation behind the order becomes clearer when you look at the specific risk it is trying to address. Several sources point to Anthropic's restricted rollout of Claude Mythos Preview as the kind of scenario regulators are worried about. That model was held back due to concerns about its autonomous vulnerability-discovery capabilities. As models become capable of meaningfully accelerating cyberattacks, the question of who knows what before release matters in a way it did not two years ago.
Critics are skeptical the order does much about that. The Atlantic called it "relatively toothless," arguing that a voluntary framework without enforcement gives the government visibility but no actual leverage. Politico acknowledged the "sea change in Washington's willingness to tighten AI oversight," while noting the order was "messy, muted and far less ambitious than Silicon Valley's critics had hoped for." The underlying concern from multiple commentators is timing: the window for meaningful AI governance may already be closing faster than any voluntary process can address.
Anthropic's position is instructive. The company has publicly called for an industry-wide slowdown or pause on frontier model development, arguing that AI systems are nearing the point where they can improve themselves without human oversight. Anthropic co-founder Jack Clark told the BBC that "right now, it's like the AI industry has a gas pedal but it doesn't have a brake pedal." The Trump order, in that framing, is not a brake; it is a speed camera that companies can choose to drive past.
The deeper tension is whether voluntary frameworks can function when the competitive pressure to release is this intense. Each lab that participates gives the government 30 days of insight; each lab that passes loses nothing and gains relative speed. Whether enough companies sign up to make the review process meaningful, or whether participation becomes a reputational signal rather than a substantive safety mechanism, will determine what this order actually amounts to. The agencies have until early August to define the rules. The answer to that question will take considerably longer.