The Trump administration is quietly reviving the pre-release AI model reviews it spent the past year mocking, after Anthropic's new Mythos model demonstrated cyber-exploit capabilities the White House considers a national security risk. Google, Microsoft, and xAI all agreed Tuesday to hand early access to new frontier models to the Commerce Department's Center for AI Standards and Innovation. That body was called the US AI Safety Institute until last June, when Commerce Secretary Howard Lutnick renamed it and declared that innovators would no longer be limited by safety standards used "under the guise of national security."
An executive order under discussion would create a working group of tech executives and government officials to formalize the review process, modeled loosely on the system Britain has built. The reversal is sharp. Vice President JD Vance told the Paris AI Action Summit in February 2025 that "the AI future is not going to be won by hand-wringing about safety," and warned that excessive rules might "kill a transformative industry just as it's taking off."
The trigger is Mythos, Anthropic's latest large language model, currently in preview with roughly 50 companies. Anthropic wants to expand that to 120. The White House is blocking the expansion on national security grounds, arguing the model is capable enough at developing cybersecurity exploits that wider access is dangerous. Officials have also raised compute-capacity concerns about Anthropic serving government customers, which Anthropic disputes.
Key facts
- 01Google, Microsoft, and xAI agreed Tuesday to give the Commerce Department's Center for AI Standards and Innovation early access to new frontier models for review.
- 02The White House is blocking Anthropic's plan to expand Mythos preview access from 50 companies to 120, citing cyber-exploit capabilities.
- 03Trump revoked Biden's AI safety-testing executive order on day one of his second term and ended the requirement three days later.
- 04One set of Trump officials is working to phase out Anthropic models across the government over six months while another expands agency access to Mythos.
- 05The Commerce Department renamed the US AI Safety Institute to the Center for AI Standards and Innovation last June.
The administration's posture toward Anthropic is now openly contradictory. Trump officials have moved to designate the company a "supply chain risk" because it refused to amend its Pentagon contract to allow "all lawful use" of its technology. One arm of the government is working to phase out Anthropic models across federal agencies over six months. Another arm is working to expand agency access to Mythos. The administration is, in effect, trying to route around its own legal roadblock.
“The White House now opposes Anthropic's plan to expand Mythos preview access from roughly 50 companies to 120, citing national security risks tied to the model's cyber-exploit capabilities.”— Jaeden Schafer
The rhetorical shift cuts against a year of dismissals. David Sacks, the outgoing White House AI and crypto czar, had labeled safety advocates a "doomer industrial complex" running a "sophisticated regulatory capture strategy based on fearmongering." Michael Kratsios, who runs the Office of Science and Technology Policy, accused international AI governance efforts of maintaining "a general atmosphere of fear." Trump revoked Biden's safety-testing executive order on the first day of his second term, then issued a replacement three days later titled "Removing Barriers to American Leadership in Artificial Intelligence" that scrapped the testing requirement.
On the state level, the administration is still pushing the other direction. Trump officials have been pressuring Republican lawmakers in Nebraska and Tennessee to weaken or kill bills that would impose safety and transparency requirements on AI companies. The federal government wants pre-release review for itself while telling state legislatures to stand down — a posture that becomes harder to defend the moment Google, Microsoft, and xAI start submitting models to Commerce.
The National Security Agency is already using Mythos to hunt vulnerabilities in Microsoft products, with the obvious implication that foreign adversaries will soon wield comparable models against US critical infrastructure if they are not already. The political ground has also shifted. Public backlash against data centers and the energy footprint of AI is putting the administration at odds with parts of its own base, and the "get rich and beat China" pitch has not traveled far beyond the venture firms that wrote it.
Critics of the new review regime are not wrong to worry about how it will be used. A licensing system controlled by the Trump White House could be deployed to deny releases of models deemed "woke," or to extract favors from companies seeking approval. The Federal Communications Commission under Brendan Carr offers a template for how a regulatory body can be turned into a political instrument. Any safety review structure built now will outlive the news cycle that produced it.
The trial pitting Elon Musk against OpenAI entered its second week as the policy reversal landed, a separate reminder of how unstable the governance of frontier AI has become. Co-founder Greg Brockman testified that Musk demanded a majority stake during for-profit conversion talks because he needed $80 billion to start a city on Mars, and threatened to launch a rival AI company with "one Tweet." Musk lawyer Steven Molo pressed Brockman on why, given a personal goal of $1 billion, he had not donated the rest of his $30 billion stake. "It takes 30 billion dollars to get you out of bed in the morning?" Molo asked.
The contradiction inside the Trump administration's AI policy is now the policy. Officials who spent a year insisting that frontier-model risks were invented by incumbents to crush competitors are quietly building the review apparatus those incumbents asked for, while still publicly attacking the people who proposed it. Anthropic is simultaneously a national security threat and a vendor the government cannot stop trying to buy from. That is what it looks like when capabilities outrun ideology — the models got more dangerous, and the people in charge noticed.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
