President Donald Trump signed an executive order Tuesday that quietly marks a shift in his administration’s approach to artificial intelligence. No public ceremony. No bold declarations from the Oval Office. Just a private act that gives federal agencies up to 30 days to review the most powerful new AI models before companies release them more widely.
The order directs the departments of Treasury, Defense, Commerce and Homeland Security to secure voluntary agreements with AI developers. It calls for classified benchmarks to assess national security implications. And it tasks Treasury Secretary Scott Bessent with coordinating an AI cybersecurity clearinghouse to scan software for flaws uncovered by these systems. Banks, utilities and hospitals sit high on the list of worried infrastructure operators.
From Hands-Off to Measured Caution
Trump had postponed a more ambitious version of this directive just weeks earlier. That draft called for a 90-day review window. Industry pushed back hard. The president worried it might slow America’s lead over China. “We’re leading China. We’re leading everybody,” he told reporters on May 21. “And I don’t want to do anything that’s going to get in the way of that lead.”
So the final product landed softer. Voluntary. Time-limited. Focused on cybersecurity threats rather than broad safety or licensing requirements. Nothing in the text authorizes mandatory preclearance. Yet it still hands government officials earlier sight of frontier models than many in Silicon Valley preferred. And it arrives at a moment when one particular model has Washington rattled.
Anthropic’s Claude Mythos changed the conversation. In April the company limited its release after the system demonstrated an extraordinary ability to discover software vulnerabilities. Executives warned of a potential “reckoning” in cybersecurity. They have since expanded access to about 150 organizations and plan further rollout of Mythos-level capabilities. The startup, which confidentially filed for an IPO, has held repeated talks with officials about these risks. (Business Insider)
But the concerns run deeper than one model. Advanced AI systems can now identify weaknesses across codebases at speeds and scales beyond human teams. Hackers or nation-state adversaries could exploit that knowledge. Banks and utility companies grew nervous enough to press the administration for action. Treasury consulted with them directly while shaping the order.
The directive builds on years of quieter voluntary testing. Companies including OpenAI, Anthropic, Google, xAI and Microsoft have already submitted models to the Commerce Department’s Center for AI Standards and Innovation. Google’s Kent Walker called the new order “an important step forward.” Anthropic and OpenAI declined immediate comment.
Trump himself framed the move in familiar terms. “Advanced AI capabilities make our Nation stronger, but also introduce new national security considerations,” the order states. “We will continue to lead an America First cybersecurity effort that enhances both our national security and our global AI dominance.” The White House released the full text hours after the private signing. (White House)
Agencies now face tight deadlines. Within 30 days, officials must issue directives to harden federal networks and critical infrastructure against AI-assisted attacks. The Pentagon gets the same window to bolster its own systems. Treasury must stand up the clearinghouse and begin collaboration on vulnerability patching. A classified benchmarking process, overseen in part by the NSA director, will help agencies judge which models qualify as high-risk.
Yet the order stops short of many tools that national security hawks had sought. No mandatory reporting. No licensing regime. No requirement that companies pause releases if reviews flag problems. Critics on the Hill and in previous administrations call it modest at best. Former officials expressed surprise that so much of the earlier language survived Trump’s cuts. One industry executive familiar with the talks described it as a tactical retreat that still gives labs breathing room.
That balance reflects months of back-and-forth. AI firms reviewed drafts. David Sacks, Trump’s AI and crypto czar, had warned against overly prescriptive rules. The final text emphasizes collaboration, intellectual property protections and accelerated government adoption of secure AI tools. It also directs the Justice Department to pursue criminal cases involving AI-enabled hacking.
Recent reporting shows the pivot gained momentum after Anthropic’s disclosures. Officials received briefings on Mythos’s capabilities. They heard from banks already running scans with AI assistance. The model didn’t target cyber offense as its primary goal. It simply excelled at it as a side effect of general training. That fact unnerved experts more than any deliberate weaponization.
So the administration chose a narrower path. Focus first on cyber. Build voluntary channels. Keep America’s innovation engine running hot. The order explicitly says nothing shall be construed to create new regulatory burdens or slow deployment unnecessarily.
Still, it sets precedents. Government now expects earlier access to the most capable systems. Agencies will develop shared benchmarks. Information on discovered vulnerabilities can flow through the new clearinghouse to defenders. Over time these pieces could evolve into something more structured.
Industry largely views the outcome as a win. The 30-day window beats the 90-day alternative. Voluntary language preserves flexibility. Companies can still move fast on releases to enterprise customers while offering government a preview. But the signal is clear. Washington no longer pretends the most powerful models pose zero risk to critical systems.
Trump’s earlier executive actions had tilted heavily toward deregulation. He revoked predecessor rules seen as stifling. He pushed to preempt state AI laws. This new directive does not reverse that stance. It carves out a targeted exception for national security cyber threats.
Executives at frontier labs have long argued self-regulation and responsible scaling suffice. Anthropic’s own decisions around Mythos lend some credence to that view. The company held back public access without government mandate. It worked with partners on controlled releases. Yet the administration decided additional visibility was warranted.
Implementation will test whether the voluntary framework delivers results. Agencies must move quickly on the 30- and 60-day requirements. Tech companies will decide how much they share and how candidly. Trust between the two sides, strained at times in recent years, now faces a practical stress test.
Broader questions linger. The order focuses on models before external release. It says less about internal experimentation at labs or risks from stolen model weights. Cyber remains the headline threat. Other national security dimensions receive less attention here.
Even so, the document marks progress for those who have urged government to treat advanced AI as a dual-use technology. It acknowledges that capabilities once confined to narrow tools now emerge as byproducts of general intelligence. And it creates mechanisms, however light, for defenders to stay ahead of the next surprise.
Trump signed the order Tuesday. By evening details had spread across policy circles and tech boards. Analysts noted the absence of fanfare. They also noted the speed with which the scaled-back version reached his desk after the May postponement.
The coming weeks will reveal whether this framework satisfies the banks and utilities still losing sleep over AI-driven vulnerability discovery. Or whether further measures become necessary. For now the administration has chosen calibrated engagement over heavy regulation. The AI race continues. Washington simply wants a better seat at the table before the next model drops.
Google welcomed the step. Others stayed silent. The clearinghouse will soon take shape. Benchmarks are in development. And Mythos-level systems keep advancing. The order may be modest. Its timing suggests the concerns have only grown sharper. (Politico) (Reuters) (Wall Street Journal)