Staff at the US office that tests AI for safety were asked to report informally to the White House, former officials say

Staff at the Commerce Department's AI testing center were recently asked to report informally to a White House office. That is according to two former government officials who spoke to the Associated Press anonymously, in a report published November 8.

For companies buying AI and the investors funding it, the stakes are practical. The center was one of the few places outside the AI companies that could publish findings on how risky a model is. Its public page lists no assessment of a model from an American developer.

It also runs on about $10 million a year, roughly one-eighth of its British counterpart's budget, according to AP. It has had no permanent director since July. The White House did not answer AP's specific questions about the center.

What changed, and what didn't

The unit is now called the Center for Advancing Innovation and Standards for Super Intelligence, or CAISSI. It began as the Biden-era US AI Safety Institute and has been renamed twice since 2025.

AP describes an informal reporting line, not a formal transfer, and did not name the White House office involved. But the move matches a plan reported in October. On October 12, Reuters carried a report from The Information that three senior CAISSI officials would move to the White House.

Those officials would report to Michael Kratsios, who runs the White House science and technology office, and National Cyber Director Sean Cairncross. A White House official cited coordination problems with Commerce. A senior Commerce official said Commerce Secretary Howard Lutnick had invited the two men to co-manage the center months earlier.

Control over what the center publishes shifted earlier still. In June, officials including Cairncross told it to pause public reports on its model reviews, the Wall Street Journal reported. The pause was to last while a new executive order took effect. Since then, the assessments on its public page have covered models from Chinese developers Z.ai and Moonshot, not US labs.

Leadership has churned. Collin Burns, a former Anthropic researcher, was reportedly pushed out four days after starting in April. His successor, Chris Fall, resigned on July 20 after about three months. The director of NIST, the center's parent agency, has served as acting head since. Commerce said in July that a permanent pick would come within weeks. None has been named.

Who now decides which AI models get a government check

The reporting change matters because of the rules that now govern pre-release testing. Executive Order 14409, signed June 2, sets up a classified process to rate how well AI models can carry out cyberattacks.

The director of the National Security Agency decides which models count as a "covered frontier model," after consulting White House and security officials. Commerce, through NIST, is only consulted. The order does not mention the testing center at all.

Developers may voluntarily give the government up to 30 days of access to such a model before releasing it to other trusted partners. The order expressly rules out mandatory licensing or government pre-approval. It requires no public disclosure, and assessments go back to developers only "as appropriate."

In short, the main government check on new AI models is voluntary, classified and run from national-security offices. Lawmakers in both parties have pushed back on the secrecy. Republican Senator Ted Budd asked in June that the center publish its research "to the extent possible." Democratic Representative Ro Khanna wrote to Lutnick in August about reports that the center was blocked from publishing model assessments.

What it means for buyers and investors

First, independent safety evidence on US models now comes mostly from the companies and the outside evaluators they hire. AP notes that companies can limit those evaluators' access and disregard their findings. Buyers doing due diligence have little public government data to check claims against.

Second, the arrangement adds political risk to technical judgments. Some critics worry that decisions on cybersecurity and bioweapon risk could be politicized. Amos Toh of the Brennan Center for Justice said such decisions should be explained in public. Logan Koepke of the nonprofit Upturn questioned what government testing adds if labs run their own. Caroline Siegel Singh of the Federation of American Scientists said policy should be shaped by nonpartisan technical experts.

Supporters see an upside. Closer ties to President Trump could give the center more prominence and possibly more money, AP reported. OpenAI has itself urged policymakers to fund it better.

Third, testing methods lag the products. There is no standard process for testing how safely AI agents, systems that carry out multi-step tasks on their own, behave in the real world, AP reported. Sayash Kapoor, an incoming assistant professor at UC Berkeley, said checking only the end result can hide serious mistakes along the way.

Three signals will show where this goes. Watch whether a permanent director is named and whether the center resumes publishing on American models. Also watch whether the White House reporting line is put in writing. Until then, the safety record of US frontier AI rests largely on what its makers choose to share.