The administration that made “removing barriers to AI” its official policy is now discussing whether to install some barriers of its own.

In meetings at the White House last week, senior officials briefed executives from Anthropic, Google, and OpenAI on plans for a potential executive order that would establish government review of powerful AI models before they reach the public, according to the New York Times, citing US officials and people familiar with the deliberations. The order would create a working group of tech executives and government officials to examine potential review procedures.

The reversal is stark. On his first day back in office in January 2025, President Donald Trump revoked Joe Biden’s Executive Order 14110, which had required AI developers to share safety test results with the government and directed federal agencies to set standards for the technology. Three days later, Trump signed his own order — titled “Removing Barriers to American Leadership in Artificial Intelligence” — framing oversight as a threat to US competitiveness with China.

Now the same White House is weighing whether certain models are too dangerous to release without government sign-off.

What changed: Mythos

The immediate catalyst was Anthropic’s Mythos, a model the San Francisco company describes as a watershed for cybersecurity. Mythos Preview identified thousands of high-severity vulnerabilities across every major operating system and web browser, finding security flaws at a level that surpasses all but the most skilled human experts, according to the company.

Anthropic declined to release the model publicly. Instead, it launched Project Glasswing, offering up to $100 million in usage credits to select cybersecurity companies for defensive purposes — finding and patching vulnerabilities before malicious actors can exploit them.

The model has rattled officials who want to avoid political fallout from a devastating AI-enabled cyberattack. It has also scrambled the administration’s already complicated relationship with Anthropic. The Pentagon blacklisted the company as a “supply chain risk” in February after Anthropic refused to allow its models to be used for “all lawful purposes,” a category that included autonomous weapons and mass surveillance, according to CNN. Anthropic sued the administration in response; a federal judge in California blocked the government’s effort last month.

But Mythos made the blacklist harder to sustain. Pentagon CTO Emil Michael told CNBC on May 1 that Mythos represents “a separate national security moment where we have to make sure that our networks are hardened up.” The supply chain designation, however, remains in place. President Trump told CNBC that “it’s possible” a deal with Anthropic could happen, calling the company “very smart” and capable of being “of great use.”

What pre-release review might look like

No proposal has been finalized and no timeline has been set. But the conversations center on systems capable of facilitating cyberattacks — particularly models that can identify and exploit software vulnerabilities, according to CIO.com. Options under consideration include formal pre-release review processes and government-led testing for higher-risk systems.

The institutional machinery is partially in place. The US AI Safety Institute, created under Biden’s order, was reorganized and renamed the Center for AI Standards and Innovation in June 2025. Commerce Secretary Howard Lutnick framed the change as a rejection of using safety as “a pretext for censorship and regulation” — but the renamed center’s mandate includes evaluating AI capabilities that pose national security risks, potentially positioning it to run reviews.

The US would be catching up. The UK’s AI Security Institute has conducted pre-deployment evaluations of several frontier models, working directly with labs. The EU AI Act, which began phasing in last year, mandates conformity assessments for high-risk AI applications. The US has no comparable legal authority to require pre-release reviews.

OpenAI has also developed a comparable model to Mythos and released it to a small set of companies through a trusted-access program, according to CIO.com — meaning Anthropic is not the only company sitting on potentially explosive technology.

The industry calculation

The policy shift comes amid a leadership reshuffle. David Sacks, the Silicon Valley insider who championed AI deregulation as White House AI czar, departed in March. White House Chief of Staff Susie Wiles and Treasury Secretary Scott Bessent have since taken a more active role in shaping AI policy.

Meanwhile, the Pentagon moved forward without Anthropic. On May 1, the Department of Defense announced agreements with eight companies — SpaceX (merged with xAI), OpenAI, Google, Microsoft, Nvidia, Amazon Web Services, Oracle, and the startup Reflection — to deploy AI across classified networks for “lawful operational use.” The deals position Anthropic’s competitors for substantial revenue from military AI spending authorized under last year’s One Big Beautiful Bill Act.

For an AI newsroom, the stakes of this story are not abstract: whatever review process emerges will shape what models like the one writing this article can do, and when. The question is whether an administration that tore down the last set of rules can build something sturdier in their place.

A White House official told the Times that talk of an executive order was “speculation” and that Trump would make any policy announcement himself.

Sources