Staff at the federal government’s main frontier AI model testing unit were recently asked to report informally to a White House office, according to an Associated Press report published Sunday, November 8, that cites two former officials who spoke anonymously. The unit, the Center for Advancing Innovation and Standards for Super Intelligence (CAISSI), is not mentioned in the order that governs federal pre-release AI testing, Executive Order 14409, which also contains no requirement to publish assessments.

The White House said President Donald Trump formally established a Super Intelligence Force and commissioned a risk report from his Council of Advisors on Science and Technology. The unit grew out of the Biden-era US AI Safety Institute, which Commerce renamed the Center for AI Standards and Innovation (CAISI) in June 2025. NIST’s own page now carries the longer CAISSI name in its title while its news items still say CAISI.

Two Former Officials Describe an Informal White House Reporting Line

The AP describes an informal arrangement, not a formal transfer. The report follows an October 12 report by The Information, relayed by Reuters, that the administration planned to move three senior CAISSI officials into the White House, where they would report to Office of Science and Technology Policy Director Michael Kratsios and National Cyber Director Sean Cairncross. According to that account, a senior Commerce official said Commerce Secretary Howard Lutnick had invited the two to co-manage CAISSI about six months earlier, and a White House official cited coordination problems with Commerce.

NIST’s page does not say who CAISSI reports to. Commerce said in July that NIST Director Arvind Raman would serve as acting CAISSI director.

Executive Order 14409 Never Mentions CAISSI and Does Not Require Publication

President Trump dated the order June 2. It directs a classified benchmarking process and lets the NSA director decide which models count as “covered frontier models,” in consultation with other officials. It also sets up a voluntary framework in which developers can give the government up to 30 days of access before releasing a covered model to other trusted partners. The text treats NIST as a consulting party, provides for sharing assessments with developers and researchers “as appropriate,” and contains no requirement to publish designations or assessments. Section 3(c) bars mandatory licensing, preclearance or permitting requirements.

The practical result is that where CAISSI reports and whether its work is public both rest on administrative discretion that the order does not constrain.

Is CAISSI Still Publishing Its AI Model Assessments?

Only in part. According to Gizmodo’s summary of a Wall Street Journal report, administration officials, including Cairncross, told the center in June to pause public reports on its model reviews while the new order was implemented. White House spokeswoman Liz Huston told the Journal that implementing the AI agenda is a “whole-of-government effort,” without directly addressing the pause.

The AP says the center has stopped publishing its research. NIST’s news page lists a narrower picture: assessments of Z.ai’s GLM-5.2 (July 17) and GLM-5.3 (September 17), a joint UK AI Security Institute assessment of Moonshot’s Kimi K3 (July 23), and a joint practices document on incident readiness (October 16). It lists no assessment of a US developer’s model among those items. In other words, public output has narrowed to Chinese models and joint guidance, and outsiders have no published CAISSI assessment of a US frontier model to check.

Lawmakers from both parties have pressed the point. Sen. Ted Budd (R-N.C.) wrote to Cairncross and Kratsios in a letter released June 30, asking that the center “to the extent possible, allow its research to be published.” Rep. Ro Khanna (D-Calif.), ranking member of the House Select Committee on the CCP, wrote to Lutnick on August 13, citing reports that CAISSI lacks permanent leadership and was blocked from publishing model assessment reports, and requested answers by September 1.

CAISSI Has Had No Permanent Director Since July

The White House reportedly pushed out Collin Burns four days after he started in April, and Commerce then named Chris Fall, who resigned after about three months, on July 20. A Commerce official told Axios that Fall’s appointment was always meant to be temporary and that Raman had been reviewing candidates; Commerce said it expected to name a permanent director within weeks. The AP reports the center still lacks a permanent leader.

Kratsios Coordinates AI Action Plan; Cairncross Said He Lacks Cyber Technical Background

The two officials named in the reporting line have documented technology-policy roles. Kratsios told a Senate subcommittee in September 2025 that his office coordinates implementation of the administration’s AI Action Plan, and he has urged Congress to set a federal framework that preempts state AI laws. Cairncross, confirmed 59-35 in August 2025, told senators he does not have a technical background in cyber and had been “on the user side.”

Critics See Politicization Risk; Some Observers See a Funding Opportunity

Critics quoted by the AP, including Amos Toh of the Brennan Center for Justice and Logan Koepke of Upturn, raised concerns that decisions touching cybersecurity and bioweapons risks could be politicized, and questioned what role government testing plays if labs also run their own evaluations. The AP also notes that some observers argue closer ties to Trump could give the division more prominence and possibly more funding.

AP Puts CAISSI Budget at $10 Million a Year, One-Eighth of UK Counterpart

The AP puts CAISSI’s budget at about $10 million a year, roughly one-eighth that of its UK counterpart. The AP adds that OpenAI has urged policymakers to give the division more money and support.

Agent Evaluations Can Miss Errors, and No Standard Test Covers Agent Safety

The testing gap is technical as well as institutional. Sayash Kapoor, an incoming assistant professor at UC Berkeley’s School of Information, told the AP that older evaluation methods check only a model’s final result, so a major mistake by an AI agent goes unseen unless evaluators examine every step it took. The AP notes there is no standard process for testing how safely agents act in the real world, even on simple tasks like booking a flight. NIST’s page lists CAISSI research posts titled “Cheating On AI Agent Evaluations” (December 2, 2025) and “Analyzing Transcripts from AI Agent Evaluations” (February 18, 2026), which show the center treats evaluation integrity as a live problem.

That is why published results matter. If assessments stay unpublished, outside researchers cannot check whether a government test examined an agent’s full sequence of actions or only its final answer. Readers who rely on frontier AI, and lawmakers who fund the unit, can ask for three things: publication of assessments of US developers’ models where security allows, a documented reporting line for CAISSI, and a public account of how agent behavior is tested.

Frequently Asked Questions

What is CAISSI, and is it the same as CAISI?

Largely yes. CAISSI is the Center for Advancing Innovation and Standards for Super Intelligence, the name NIST’s page title uses for the Commerce unit that its news items still call CAISI, the Center for AI Standards and Innovation. It descends from the US AI Safety Institute.

Does Executive Order 14409 require CAISSI to publish its AI model assessments?

No. The order does not mention CAISSI, assigns the covered-model designation to the NSA director, and contains no public-reporting requirement, according to Executive Order 14409’s text.

Who runs CAISSI right now?

Commerce said NIST Director Arvind Raman would serve as acting director after Chris Fall resigned on July 20, and the AP reports no permanent director has been named.

Why is testing AI agents harder than testing chatbots?

Agents take actions in the real world, so checking only a final answer can miss serious errors along the way, and the AP reports there is no standard process for testing agent safety.