Staff at the small U.S. government office responsible for testing the safety of frontier AI models have recently been asked to report informally to a White House office, the Associated Press reported on Saturday, prompting fresh concerns about the independence of the country's top AI safety evaluator.
The office, the Center for Advancing Innovation and Standards for Super Intelligence — known as CAISSI and housed inside the Commerce Department — has also stopped publishing its research, according to the AP. The report lands amid growing public anxiety over a string of stories about out-of-control AI agents, and it describes a watchdog that is under-staffed, leaderless and increasingly quiet. For more context on this story, see our ongoing breaking AI news.
What CAISSI Does — and What the AP Found
CAISSI is charged with assessing the safety of new AI offerings before they reach the public, a mandate that includes evaluating how well frontier models resist misuse for cyberattacks and other serious harms. According to the AP, staff at the center were "recently" asked to report informally to an office of the White House. The news agency cited two former government officials with knowledge of the discussions, who spoke anonymously to protect current employees. The White House did not respond to specific questions about the center.
The scale of the operation is modest. The AP reports that CAISSI receives about $10 million a year — roughly one-eighth of what its United Kingdom counterpart receives, even though the United States is where most frontier models are built.
The informal reporting arrangement is only the latest disruption at the center. The AP describes a leadership vacuum stretching back months: a director was hired in April and fired within days, his replacement lasted a few months before resigning, and the head of the parent institute is currently serving as acting director. In practice, that means nobody with a permanent mandate is currently defending the center's independence.
A Watchdog That Stopped Publishing
The more consequential change, according to the AP, is what the center is no longer doing. In August, Congress asked why CAISSI had stopped publicly releasing reports on how AI systems perform on safety tests. Pre-deployment testing functions as a public check on model builders only if its results are visible outside the lab — and, for now, they are not.
The shift did not come out of nowhere. In October, Reuters, citing a report from The Information, revealed that the administration planned to move three senior CAISSI officials to the White House. A White House official told The Information at the time that coordination problems with Commerce showed the need to bring the center's leadership closer to AI decision-making. According to that report, the officials would answer to Michael Kratsios, who runs the Office of Science and Technology Policy, and National Cyber Director Sean Cairncross. Commerce Secretary Howard Lutnick had invited the two to co-manage the center roughly six months earlier.
Read together, the AP's new reporting suggests that plan is arriving in practice — informally, and ahead of any formal restructuring.
Independence Concerns, in Both Directions
Critics quoted by the AP worry that decisions touching cybersecurity and bioweapons risk could be politicized, or that findings could be softened to suit the technology industry. Amos Toh of the Brennan Center for Justice told the AP that such decisions are of immense public consequence and should be explained in public. Caroline Siegel Singh of the Federation of American Scientists described centers like CAISSI as powerful institutions that can drastically improve lives or do the opposite.
There is a counterargument, and the AP does not dismiss it. Evaluating a model's ability to help write malware is a national security question, and an administration that wants visibility into who sees those results has a defensible case. The difficulty is institutional: a referee that answers to one of the teams' sponsors is poorly placed to call the game straight, and the center has stopped publishing the reports that would otherwise demonstrate its independence.
Why It Matters for Builders and Buyers
For companies building on frontier models, the practical risk identified in the reporting is not a ban — it is uncertainty. If evaluations remain secret and feed into export controls and government procurement, a lab could lose a contract or a market on the basis of a finding it never gets to see or contest. That risk would propagate through the supply chain to downstream products without any warning.
Two things are worth watching next. The first is whether the administration names a permanent director for the center, and who picks them. The second is whether Congress, which has already pressed the center over its silence on published reports, moves to write the center's independence into law. Until either happens, America's frontier AI safety referee operates with a $10 million budget, no permanent leader and a reporting line that runs toward the White House.
---
Stay Ahead of AIGet the latest AI news, analysis, and breakthroughs — all in one place.
Read more AI news →
