In the race for AI, every day counts.
The order establishes voluntary AI model review before public launch, coordinating Treasury, NSA, and CISA to identify security vulnerabilities in critical infrastructure systems. Major AI companies including OpenAI, Google, and Anthropic agreed to the framework after a concerning incident where Anthropic's Mythos model exposed banking and government system vulnerabilities.
- Trump signed an executive order establishing voluntary AI model review before public launch
- Review timeline compressed from 90 days to 30 days
- Anthropic's Mythos model exposed vulnerabilities in banking, government, and hospital systems in early 2026
- OpenAI, Google, and Anthropic agreed to participate in the framework
Trump signed an executive order granting the U.S. government early access to AI models through voluntary submission by companies like OpenAI and Google, mirroring Biden's approach but with a 30-day review timeline instead of 90 days.
President Trump signed an executive order this week that grants the federal government early access to artificial intelligence models before they reach the public. The arrangement is voluntary—companies submit their systems for review rather than face mandatory government screening—but it represents a striking reversal for an administration that had, until recently, resisted any regulatory framework for AI development in the name of staying competitive with China.
The order emerged from negotiations with the country's leading AI firms: OpenAI, Google, and Anthropic. Under the new rules, these companies can choose to have their models evaluated by federal agencies before launch. The Treasury Department, the National Security Agency, and CISA will coordinate a new center focused on digital security in AI systems, working alongside private companies and operators of critical infrastructure to spot vulnerabilities and prioritize fixes. The language is careful to note that this does not constitute mandatory prior government control—the emphasis on voluntariness matters politically, even if the practical effect is that major firms have already agreed to participate.
The timing of this order reflects a specific anxiety that gripped Washington earlier this year. In the spring, Anthropic developed a model called Mythos that demonstrated an unsettling capability: it could expose security flaws in digital systems across banking, government, and hospitals. The company chose not to release it publicly. That incident crystallized concerns among policymakers that the race to deploy ever-more-powerful AI systems might outpace the ability to understand their risks. The voluntary review framework is, in some sense, a response to that moment of recognition.
What makes this order notable is how much it echoes the approach of Trump's predecessor. In 2023, Joe Biden issued an executive order requiring companies to share the results of their safety testing. Trump revoked that measure when he returned to office, calling it too restrictive and worried it would hamper American advantage over China. Yet the order signed this week is structurally similar to Biden's framework—the main difference being a compressed timeline. Companies now have 30 days to submit their models for review, down from the 90 days Biden's order had proposed. "In the race for AI, every day counts," David Sacks, a White House adviser on AI policy, said in explaining the acceleration.
The reversal did not come smoothly. An earlier version of the order was scheduled to be signed on May 25, but Trump canceled it just hours before the ceremony, saying he disagreed with certain aspects and did not want to "compromise" American advantage. Analysts pointed to Sacks as the influential voice that persuaded the president to reconsider and move forward with a modified version. That episode exposed real tension within the administration between those who see some regulatory guardrails as necessary and those who view any government oversight as a drag on innovation.
Sam Altman, the chief executive of OpenAI, has already signaled his opposition. He plans to speak against the measure during a visit to Washington this week and will ask Congress to increase funding for AI testing at the Commerce Department. His resistance suggests that even with the voluntary framing and the compressed timeline, the tech industry sees this as a meaningful constraint—or at least wants to negotiate for more resources to comply with it. The order is now in effect, but the conversation between government and industry about what AI oversight should look like is far from settled.
Citações Notáveis
In the race for AI, every day counts.— David Sacks, White House adviser on AI policy
A step important that offers cybersecurity defenders more tools to stop malicious actors.— Kent Walker, Google's public affairs lead
A Conversa do Hearth Outra perspectiva sobre a história
Why did Trump reverse course so quickly on this? He seemed genuinely opposed just days before.
The May 25 cancellation was real resistance—he worried the order would tie America's hands against China. But Sacks and others convinced him that some framework was necessary after the Mythos incident. The revised version compressed the timeline to 30 days, which was a concession to speed.
So the companies agreed to this voluntarily? That seems almost too neat.
They did agree, but "voluntary" is a careful word. These are the companies that will dominate AI for years. Participating in a government review process is now the cost of being taken seriously. Refusing would invite much harsher regulation.
And Altman's opposition—is that genuine or theater?
Probably both. He genuinely believes the government should fund more testing infrastructure rather than just review what companies build. But opposing it also signals to his board and investors that he's fighting for their interests. The real negotiation is just beginning.
What happens if a company refuses to submit a model?
The order doesn't say. That ambiguity is intentional. The threat of what *could* happen—mandatory review, public disclosure of vulnerabilities, congressional hearings—is probably enough to keep companies compliant without explicit enforcement.
So this is less about security and more about control?
It's both. The Mythos incident was real—those vulnerabilities in banking systems are genuinely dangerous. But the order also lets the government see what's coming before it's public, which is a form of control, yes. The companies accepted it because the alternative looked worse.