The Associated Press reported on November 8, 2026, that officials at the Center for Advancing Innovation and Standards for Super Intelligence, or CAISSI, had recently been asked to report informally to a White House office, raising concerns about the independence of federal AI safety evaluations.
Two former government officials familiar with ongoing conversations described the request on condition of anonymity to avoid repercussions for current staffers. The office was not identified, and the reported discussions concern an informal arrangement rather than an announced formal reorganization.
Critics warned that the arrangement could politicize decisions affecting cybersecurity and bioweapons, with findings potentially altered or muted in deference to a technology industry largely entrusted to police itself. Caroline Siegel Singh, an AI and emerging technology program manager at the Federation of American Scientists, argued for nonpartisan or bipartisan technocratic experts to make policy in the public interest.
Others argued that closer ties to President Donald Trump could give the division greater prominence and possibly more funding.
Leadership Turnover and Public Output
CAISSI is an independent Commerce Department division and until recently was called the U.S. Center for AI Standards and Innovation. It is allocated $10 million annually, about one-eighth the annual budget of its counterpart in the United Kingdom.
Its director was appointed in April 2026 and fired within days. His replacement resigned after a few months, leaving the center without a permanent leader; the head of its parent institute also serves as its acting director.
Congress began asking in August 2026 why the center had stopped releasing reports on how safely certain AI systems were performing. OpenAI had also urged policymakers to provide the division with more funding and support.
NIST’s official CAISSI page provides a dated record of different kinds of output. Its most recent listed research-blog entry is dated March 23, 2026, while its news section lists a September 17, 2026 assessment of Z.ai’s GLM-5.3 cyber capabilities and an October 16, 2026 publication on incident readiness and response.
The Existing Voluntary Framework
NIST describes CAISSI as industry’s primary government contact for testing and collaborative research involving commercial super-intelligence systems. Its mandate includes developing measurement guidelines, best practices and voluntary standards, establishing voluntary agreements with developers and evaluators, and leading unclassified evaluations of demonstrable national-security risks involving cybersecurity, biosecurity and chemical weapons.
The center is also assigned to assess adversary systems for vulnerabilities, backdoors and other covert malicious behavior. Its mandate includes coordinating evaluation methods and assessments with the Defense and Energy departments, the Department of Homeland Security, the Office of Science and Technology Policy, and the Intelligence Community.
Executive Order 14409, issued June 2, 2026, directs national-security agencies, in consultation with NIST and others, to develop a classified process for benchmarking advanced cyber capabilities. That process is intended to establish the threshold for designating a “covered frontier model.”
The order also calls for a voluntary framework under which developers could provide government access to covered frontier models for up to 30 days before their planned release to other trusted partners. It expressly states that it does not authorize mandatory government licensing, preclearance or permitting for developing, publishing, releasing or distributing AI models.
The independence dispute therefore concerns who directs and communicates technical findings within that voluntary system, rather than a new federal power to approve models before release.
Company-Controlled Evaluations
Trump asked leading AI companies in September 2026 to undertake extensive self-policing through a voluntary safety pact. The administration has not explained how it will monitor implementation.
The evaluation problem is also changing as models are deployed as AI agents that can plan and execute multi-step tasks without continual human prompting. Sayash Kapoor, an incoming assistant professor at the University of California, Berkeley School of Information, said there is no standard process for testing how safely agents act in real-world settings. He used the example of an agent booking a flight: checking only the final result can make the task appear successful. “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 said.
Private evaluations do not necessarily supply an independent backstop. Anthropic CEO Dario Amodei and OpenAI CEO Sam Altman have promoted independent evaluation firms, but their companies can restrict those firms’ access and set aside their evaluations.
The White House did not respond to the specific questions about CAISSI, but said Trump had formally established a Super Intelligence Force and commissioned a risk report from the Council of Advisors on Science and Technology.
Logan Koepke, a senior project director at the civil rights and technology nonprofit Upturn, said company-run internal evaluations and company-hired external reviewers leave the role of government testing unclear.
Amos Toh, senior counsel at the Brennan Center for Justice, said allowing laboratories to police their own models without clarifying the government’s role weakens independent checks on systems that can cause real-world harms, and that consequential decisions should be explained and justified publicly.
The reported request leaves a concrete governance question: who sets CAISSI’s evaluation priorities, and how visible its methods and findings will be. Closer White House ties may alter its prominence, but they do not by themselves resolve how independent government testing will function alongside voluntary developer access and company-controlled private reviews.
Credit: Source link
