Product1 publisher3 min readPublished
Trump's executive order settles on a voluntary 30-day preview of frontier models
Fast Company reports that Treasury secretary Scott Bessent backed a 90-day government preview of new frontier models so agencies could patch first. The order that emerged asks for 30 days, voluntarily, and deployers read whatever the vendor chooses to publish.
The Product Desk · Product desk
What happened
- Trump revoked Biden's AI executive order on his first day back in office, ending the requirement that developers of the most powerful systems share certain safety-testing results with the federal government.
- The executive order that emerged set a voluntary 30-day window instead, after Trump adviser David Sacks argued that 90 days would be too burdensome.
- Anthropic CEO Dario Amodei published an essay on Saturday calling for slower model development, third-party audits including embedded researchers, and international coordination on safety.
Compiled by The Product DeskSomething wrong?How this is made
Why it matters
- exposure The 90 days were argued for so federal defense and intelligence systems could be patched before release; a company running the same widely used software gets whatever notice the lab volunteers.
- decision A buyer who needs a safety-testing artifact for internal sign-off has to purchase it as a contract term, since no federal filing will produce one.
- cost With AI investment tied to roughly a third of this year's U.S. growth on ING's estimate, any mandated slowdown gets priced as a macroeconomic cost before it is weighed as a safety measure.
- contradiction The president calls the safety warnings a hoax while his Treasury secretary prepares bilateral AI talks with China in which the U.S. has floated labs on both sides monitoring themselves.
Consider the person who signs the security review before a frontier model goes into a production workflow. What goes in the file is whatever the lab decided to publish. Anthropic's May report is the working example: the company said Mythos, then in preview with select organizations, had shown a remarkable ability to quickly detect and exploit vulnerabilities in widely used software [3]. Anthropic published that report on its own. It is also what rattled administration officials worried about national security, Bessent among them [4].
The government's own ask had a number on it. Bessent backed giving it a 90-day preview of new frontier models before public release, enough time to shore up systems including those used by defense and intelligence agencies, according to the Financial Times and Fast Company's Mark Sullivan [5]. The executive order that emerged set 30 days, voluntary, after David Sacks argued that 90 would be too burdensome [6]. A third of the window Bessent asked for, and voluntary [15].
Trump has described the safety scare as a hoax meant to slow the U.S. in its competition with China, and has rejected recent calls from AI executives and researchers for more oversight [8]. On Monday he wrote on Truth Social: "[T]he only control or 'guardrails' that AI needs is a STRONG AND SMART (High IQ!) PRESIDENT, and the U.S.A. has that, in spades!" [9] On Saturday, Anthropic CEO Dario Amodei published an essay asking for the opposite: slower model development, third-party audits including embedded researchers, and international coordination on safety [7].
The Biden requirement went away in January 2025, when the most capable systems were Anthropic's Claude 3.5 Sonnet and OpenAI's GPT-4o [1][2]. Fast Company reports that by June 2026 the strain of fitting the hands-off posture to more capable models was beginning to show [14]. Between those two dates sit roughly 17 months in which no federal rule required a developer to hand over safety-testing results [16].
There is a second reading of why the ask got smaller. Sacks, a science and technology adviser to Trump, and other accelerationists on X have floated that the big labs raise existential risk in part hoping to frighten the government into regulations that entrench market leaders and disadvantage smaller players, including open-source developers [12].
For a team shipping on top of these models, assurance comes from one of three places: a document the vendor chose to publish, a term negotiated into the contract, or a test run in-house. A federal filing is not on that list, and Sullivan does not address state AI laws or procurement rules [17]. The document is free and arrives when the vendor decides. Negotiating the contract term takes leverage a small buyer lacks. An in-house test costs staff. If the model will touch software you also operate, I would spend on the in-house test and accept that you are duplicating work the lab already did, because the preview Bessent wanted was meant to protect government systems [5].
What to watch
- Whether any lab actually uses the voluntary 30-day window on a frontier release, and says publicly that it did.
- Whether Bessent's bilateral AI talks with China produce any self-monitoring or information-sharing commitment from labs on either side.
- Whether Mythos ships beyond preview, and what its release documentation says about vulnerability discovery.