The Associated Press reported on November 8, 2026, that officers on the Middle for Advancing Innovation and Requirements for Tremendous Intelligence, or CAISSI, had lately been requested to report informally to a White Home workplace, elevating issues in regards to the independence of federal AI security evaluations.
Two former authorities officers accustomed to ongoing conversations described the request on situation of anonymity to keep away from repercussions for present staffers. The workplace was not recognized, and the reported discussions concern a casual association moderately than an introduced formal reorganization.
Critics warned that the association may politicize choices affecting cybersecurity and bioweapons, with findings doubtlessly altered or muted in deference to a know-how trade largely entrusted to police itself. Caroline Siegel Singh, an AI and rising know-how program supervisor on the Federation of American Scientists, argued for nonpartisan or bipartisan technocratic specialists to make coverage within the public curiosity.
Others argued that nearer ties to President Donald Trump may give the division better prominence and probably extra funding.
Management Turnover and Public Output
CAISSI is an impartial Commerce Division division and till lately was known as the U.S. Middle for AI Requirements and Innovation. It’s allotted $10 million yearly, about one-eighth the annual price range of its counterpart in the UK.
Its director was appointed in April 2026 and fired inside days. His alternative resigned after just a few months, leaving the middle with out a everlasting chief; the top of its dad or mum institute additionally serves as its performing director.
Congress started asking in August 2026 why the middle had stopped releasing studies on how safely sure AI methods had been performing. OpenAI had additionally urged policymakers to supply the division with extra funding and assist.
NIST’s official CAISSI page gives a dated document of various sorts of output. Its most up-to-date listed research-blog entry is dated March 23, 2026, whereas its information part lists a September 17, 2026 evaluation of Z.ai’s GLM-5.3 cyber capabilities and an October 16, 2026 publication on incident readiness and response.
The Current Voluntary Framework
NIST describes CAISSI as trade’s major authorities contact for testing and collaborative analysis involving industrial super-intelligence methods. Its mandate consists of growing measurement tips, finest practices and voluntary requirements, establishing voluntary agreements with builders and evaluators, and main unclassified evaluations of demonstrable national-security dangers involving cybersecurity, biosecurity and chemical weapons.
The middle can also be assigned to evaluate adversary methods for vulnerabilities, backdoors and different covert malicious habits. Its mandate consists of coordinating analysis strategies and assessments with the Protection and Power departments, the Division of Homeland Safety, the Workplace of Science and Expertise Coverage, and the Intelligence Group.
Executive Order 14409, issued June 2, 2026, directs national-security businesses, in session with NIST and others, to develop a labeled course of for benchmarking superior cyber capabilities. That course of is meant to determine the brink for designating a “covered frontier model.”
The order additionally requires a voluntary framework beneath which builders may present authorities entry to coated frontier fashions for as much as 30 days earlier than their deliberate launch to different trusted companions. It expressly states that it doesn’t authorize obligatory authorities licensing, preclearance or allowing for growing, publishing, releasing or distributing AI fashions.
The independence dispute subsequently issues who directs and communicates technical findings inside that voluntary system, moderately than a brand new federal energy to approve fashions earlier than launch.
Firm-Managed Evaluations
Trump requested main AI corporations in September 2026 to undertake in depth self-policing by means of a voluntary security pact. The administration has not defined the way it will monitor implementation.
The analysis drawback can also be altering as fashions are deployed as AI brokers that may plan and execute multi-step duties with out continuous human prompting. Sayash Kapoor, an incoming assistant professor on the College of California, Berkeley College of Info, mentioned there isn’t a normal course of for testing how safely brokers act in real-world settings. He used the instance of an agent reserving a flight: checking solely the ultimate consequence could make the duty seem profitable. “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,” he mentioned.
Personal evaluations don’t essentially provide an impartial backstop. Anthropic CEO Dario Amodei and OpenAI CEO Sam Altman have promoted impartial analysis corporations, however their corporations can prohibit these corporations’ entry and put aside their evaluations.
The White Home didn’t reply to the particular questions on CAISSI, however mentioned Trump had formally established a Super Intelligence Force and commissioned a danger report from the Council of Advisors on Science and Expertise.
Logan Koepke, a senior venture director on the civil rights and know-how nonprofit Upturn, mentioned company-run inside evaluations and company-hired exterior reviewers go away the function of presidency testing unclear.
Amos Toh, senior counsel on the Brennan Middle for Justice, mentioned permitting laboratories to police their very own fashions with out clarifying the federal government’s function weakens impartial checks on methods that may trigger real-world harms, and that consequential choices needs to be defined and justified publicly.
The reported request leaves a concrete governance query: who units CAISSI’s analysis priorities, and the way seen its strategies and findings shall be. Nearer White Home ties might alter its prominence, however they don’t by themselves resolve how impartial authorities testing will perform alongside voluntary developer entry and company-controlled personal opinions.

