US AI safety staff were asked to report informally to White House, AP says

Staff at the federal center responsible for assessing AI safety were recently asked to report informally to a White House office, according to the Associated Press. The request has raised concerns about political influence over technical judgments at a division that already lacks a permanent leader and has stopped publishing its research.

The Center for Advancing Innovation and Standards for Super Intelligence, or CAISSI, is an independent division of the Commerce Department. AP based its account on two former government officials familiar with ongoing conversations. They spoke anonymously to avoid repercussions for current staff. The request concerns informal reporting, rather than an announced formal transfer of the center.

The White House did not answer AP’s specific questions about CAISSI. Critics interviewed by the news agency worry that closer executive oversight could politicize decisions involving cybersecurity and biological weapons, or allow findings to be altered or muted in deference to technology companies.

Leadership turnover and unpublished findings

The reporting request comes after a succession of leadership disruptions. In April, the division’s director was hired and fired within days. His replacement resigned after a few months. The center has remained without a permanent leader since then, with the head of its parent institute also serving as acting director.

The center’s retreat from public research has drawn congressional scrutiny. In August, Congress began asking why it had stopped releasing reports on AI systems’ safety. OpenAI had also urged policymakers to provide the division with more funding and support. Until recently, the organization was called the U.S. Center for AI Standards and Innovation.

Testing more than the final answer

Federal AI safety testing has depended on voluntary cooperation from developers. Under President Joe Biden, leading companies agreed to share details with the new Commerce division and follow administration guidelines for testing and transparency reports. Government evaluators thus had a role alongside the companies building the systems, rather than leaving assessments entirely to the labs.

In September, President Donald Trump asked leading AI companies to undertake “tremendous self-policing” through a voluntary safety pact. His administration has not explained how it will monitor implementation. Meanwhile, testing faces a harder task as AI agents can plan and carry out multiple steps without human prompting.

There is still no standard process for testing agents’ safety in real-world activity, said Sayash Kapoor, an incoming assistant professor at UC Berkeley’s School of Information. He used flight booking to explain the problem: checking whether a reservation was completed correctly can miss mistakes made along the way.

But unless you’re looking at the entire set of steps that an agent took, you won’t be able to tell that the agent made a huge mistake.

Sayash Kapoor, speaking to the Associated Press

Closer to power, but on whose terms?

Not everyone sees proximity to the White House as a loss. AP also described the argument that bringing safety officials closer to the president could give the division greater prominence and perhaps more funding. Those are potential benefits, not announced commitments accompanying the reporting request.

Anthropic and OpenAI leaders have promoted independent evaluation firms, but developers can limit those firms’ access or reject their evaluations. Logan Koepke, a senior project director at the nonprofit Upturn, questioned what role government evaluators would have if companies are encouraged to build internal testing systems or hire outside assessors.

For Amos Toh, senior counsel at the Brennan Center for Justice, leaving labs to police themselves while failing to clarify government’s testing role weakens independent checks. He argued that decisions with major public consequences should be explained publicly—a transparency demand that extends beyond who receives the center’s internal reports.