The only federal office that tests frontier AI models may no longer answer only through the Commerce Department.

On 8 November the Associated Press reported that officials at the Center for Advancing Innovation and Standards for Super Intelligence (CAISSI) — the NIST unit until recently called the Center for AI Standards and Innovation, and before that the US AI Safety Institute — "recently were asked to informally report to an office of the White House." The report rests on two former government officials, who spoke anonymously to avoid repercussions for current staff. The AP does not name the office, and the White House did not answer its specific questions, pointing instead to President Trump's new Super Intelligence Force and a risk report commissioned from his Council of Advisors on Science and Technology.

The reporting line is informal, which is exactly what makes it hard to verify and easy to wave away. No document has been published.

The staffing record is not in doubt. In a letter dated 13 August, Rep. Ro Khanna, ranking member of the House Select Committee on the CCP, told Commerce Secretary Howard Lutnick that the centre's April director, Collin Burns, was "ousted just four days into his tenure," and that "in July, CAISI Director Chris Fall resigned only three months after assuming his role." The NIST director, Arvind Raman, has been acting head since, according to the letter; the AP describes the same sequence.

Khanna's letter asks about the centre's "recent cessation of publishing information on PRC AI models' propaganda and censorship" and cites public reports that it "was blocked from publishing model assessment reports." It set a 1 September deadline for answers; no reply has been published. The AP's summary — that the centre "has stopped publishing its research" — is broader than NIST's own record. Since July, NIST lists a joint assessment with the UK of the Chinese model Kimi K3's cyber capabilities (23 July, updated 28 August) and, on 16 October, ten voluntary practices for handling incidents during AI training and evaluation, published with the NSA and CISA. Neither is a model assessment of the kind Khanna asked about.

Why it matters: the unit writes voluntary standards and has no power to block a model's release. Its budget is about $10 million a year, roughly an eighth of the UK AI Security Institute's. Companies increasingly pick their own evaluators; in September Trump asked AI companies for "tremendous self-policing" under a voluntary accord without saying how compliance would be checked. Last week METR, one of those private evaluation firms, said Anthropic redacted most of the evidence it found most informative. Government testing was supposed to be the check that does not depend on a company's consent.

The political backdrop does not simplify things. In an interview with CBS's 60 Minutes aired on 8 November, Trump told Ross Douthat that AI risk "can be managed" and that he doesn't "mind guardrails," while warning that rules must not leave the United States in "a deep second position" behind China. Critics read the informal reporting line as a way to shape what the testers say in public; supporters read it as proximity that brings prominence and money. Neither reading has been confirmed.

What is confirmed is narrower and still notable: a testing body with no permanent director, a congressional request for answers that has gone unanswered in public, and, according to two former officials, a second, informal channel to the White House.