President Trump signed an executive order Tuesday establishing a pre-release testing framework for frontier AI models, reversing course after initially canceling the draft last week over concerns it would be "too burdensome." The final version reduces the government's advance access window from 90 days to 30 days—a concession that critics say is cosmetic rather than substantive. The order establishes what's being called a 'voluntary framework' for cybersecurity benchmarking of covered frontier models, with agencies directed to coordinate within two months on implementation details.

What's Actually in Section 3

The meat of the order—Section 3, titled 'Secure Frontier Model Development'—calls for a classified benchmarking process to determine which AI systems qualify as "covered frontier models." Participating labs must provide government agencies and designated 'trusted partners' with early access to upcoming releases for up to 30 days before those models reach other stakeholders. The order explicitly states it does not create any mandatory licensing, preclearance, or permitting requirement—but the air quotes around that claim are doing heavy lifting, according to analysts tracking the policy. "Make no mistake—this is a de facto mandatory governmental licensing, preclearance and permitting requirement," wrote commentator Dean W. Ball in his analysis of the order. "There is a reason 'voluntary framework' gets put in air quotes." The reasoning is straightforward: any lab developing frontier-class AI systems that wants to work with the federal government, attract government-adjacent investors, or avoid becoming a regulatory target will effectively have no choice but to participate.

Industry Embraces the Order

Microsoft President Brad Smith offered measured approval, calling it "an important step toward advancing innovation while protecting the security of the American public." Anthropic issued a similar statement, saying the order is "an important step in strengthening America's leadership in AI" and pledging to collaborate with the White House on implementation. Notably absent from the official corporate response: OpenAI, Google DeepMind, Meta AI, and other major frontier labs have not publicly commented on their positions.

The Sacks Flip-Flop

Perhaps most striking is the reversal of David Sacks, the tech investor and informal AI advisor who had successfully lobbied to delay and nearly kill the original draft. When Trump announced he wouldn't sign the initial version because it was too regulatory, Sacks appeared to have won—until the administration signed essentially the same order with one modification. Sacks then pivoted hard, praising the 30-day window as "a game changer" while insisting the EO would not create an "FDA for AI." Critics noted the irony of his previous warnings about other proposed regulations that would supposedly cripple innovation—warnings he no longer applies to this framework. Samuel Hammond offered a more sympathetic read: "I'm glad this EO exists and am less concerned about predeployment review mutating into a licensing regime than Dean," though he shared concerns about transparency.

Classified Thresholds, Opaque Process

One of the most contentious elements is that the specific regulatory thresholds—exactly what computational or capability benchmarks trigger the pre-release review requirements—are classified. This means AI researchers inside frontier labs may not know whether their current training runs are subject to the order until it's too late. "Most lab staff don't have clearances, but if the literal regulatory thresholds that trigger pre-deployment review are classified, researchers themselves won't know whether what they are training is regulated by this EO," Ball observed. Dean Ball's prediction: "My bet is they're classifying the benchmarking process to hide the fact that they're not going to be able to agree to a regulatory threshold better than 10^26 flops." The order also places NSA and intelligence community agencies at the center of implementation rather than civilian bodies like NIST's CAISI, which some experts argue would offer more transparency and less mission creep risk.

What Happens Next

Agencies have two months to coordinate on implementation details, including what constitutes a covered frontier model, how the benchmarking process will work, and who qualifies as trusted partners for early access. The order explicitly calls for this framework to be replaced by congressionally-passed law—something advocates see as critical given the fragility of executive orders across administrations. "We should try to turn this into an actual law as soon as possible," the source analysis concludes. "The next President in 2029 could reverse course with a penstroke, for better or worse." Congress has already shown interest: Representative Lori Trahan outlined competing legislative priorities including mandatory safety audits, whistleblower protections, and updated WARN Act requirements for AI-driven layoffs.

Key Takeaways

  • The EO establishes 'voluntary' pre-release review with up to 30 days government access before frontier model releases
  • Critics say the framework is effectively mandatory despite its stated voluntary status
  • Regulatory thresholds are classified, raising transparency and due process concerns
  • NSA/spy agencies lead implementation rather than civilian bodies like CAISI
  • The order tees up future licensing regimes and represents a win for safety-oriented factions over accelerationists