The U.S. government office responsible for testing advanced AI systems has been pulled closer to the White House while its public reporting has narrowed. Officials at the Center for Advancing Innovation and Standards for Super Intelligence, or CAISSI, were recently asked to report informally to an unidentified White House office, the Associated Press reported Sunday, citing two former government officials familiar with the discussions.
The center also lacks a permanent director and operates on about $10 million a year, according to the report. The White House did not answer the AP’s specific questions about the arrangement. It instead pointed to President Donald Trump’s newly established Super Intelligence Force and a separate risk review by the President’s Council of Advisors on Science and Technology.
That combination of informal oversight, unstable leadership and selective publication matters because CAISSI is the federal government’s main civilian evaluator of frontier models. It can test systems before or after release, develop benchmarks that companies cannot train against, and compare vendor claims using the same harnesses and task budgets. Those functions are most valuable when researchers, buyers and policymakers can inspect enough of the methods and results to understand what the tests actually show.
CAISSI has not stopped publishing everything
The public record is more complicated than saying the center has gone completely dark. Its NIST landing page lists an October 16 incident-readiness guide and a September 17 assessment of Z.ai’s GLM-5.3. It also published a joint cyber evaluation of Moonshot AI’s Kimi K3 with the U.K. AI Security Institute in July. The center’s research blog, however, shows no new entry after March 23.
The visible work has also become heavily concentrated on cyber capability. In the GLM-5.3 assessment, CAISSI tested vulnerability discovery and exploit development across four benchmarks. GLM-5.3 solved 74 of 183 SEC-Bench Pro tasks and reached a 61.1% ExploitBench score, while the undisclosed best U.S. frontier model reached 165 of 183 and 100%, respectively. CAISSI estimated that the Chinese open-weight model trailed the U.S. cyber frontier by about four months.
Those are unusually concrete results. The test description identifies the agent harness, turn limits, maximum reasoning settings and four task sets, including 297 private OSS-Fuzz problems. It also explains the statistical model used to combine task performance into a capability index. That level of disclosure lets outside researchers challenge the setup and prevents a headline number from floating free of its methodology.
But the public pages do not resolve the wider concern. In August, Rep. Ro Khanna asked the Commerce Department why CAISI had stopped publishing findings about propaganda and censorship in Chinese AI models. His letter also sought explanations for leadership removals, deleted documentation about model-testing agreements and reports that assessments had been blocked from publication. Sen. Ted Budd, a Republican, had made a similar request in June, arguing that public findings support model and benchmark developers.
The center’s mandate is broader than cyber benchmarks
NIST describes CAISSI as the government’s primary industry contact for testing commercial advanced-AI systems. Its published mandate includes unclassified evaluations of cybersecurity, biosecurity and chemical-weapons capabilities; assessments of foreign systems and covert behavior; voluntary agreements with model developers; and coordination with the Defense, Energy and Homeland Security departments, the intelligence community and the White House science office.
That remit reaches beyond measuring whether a model can find a browser flaw. A government evaluator can examine safety controls that a model provider would not voluntarily stress in public, compare prerelease systems under consistent conditions and retain private test sets that are less vulnerable to benchmark contamination. It can also publish negative findings that a vendor has little commercial incentive to emphasize.
The June executive order on AI cybersecurity complicated that role. It assigned the NSA, CISA and Treasury responsibility for a classified cyber benchmark and gave the NSA director a leading role in determining when a system qualifies as a covered frontier model. CAISI was not given a named role in that process, as TechsCurrent reported when the order was issued. A classified track may be necessary for sensitive attack capabilities, but it does not answer who is responsible for producing evidence that software makers, independent researchers and the public can inspect.
The rename changed the label, not the testing problem
CAISSI is the latest name for the office originally created as the U.S. AI Safety Institute and later renamed the Center for AI Standards and Innovation, or CAISI. A September 29 executive order directed federal agencies to replace the terms “artificial intelligence” and “AI” with “Super Intelligence” and “SI” in official communications. NIST subsequently expanded the center’s acronym by another letter.
The order itself says the new term covers the same technologies already included in the federal statutory definition of artificial intelligence. The rebrand therefore did not turn current models into superintelligence or alter the basic measurement work. It did make the center’s mission harder to follow at the moment when continuity and clear public records are most useful.
The White House’s Super Intelligence Force adds another layer. According to the AP, the administration has framed the group around national security, including cyber and biological risks. CAISSI already works in those domains, so the practical questions are about authority: which office chooses the test agenda, who approves publication, and whether unfavorable findings can be delayed or withheld.
Private evaluators cannot replace public accountability
AI companies increasingly hire outside evaluators and publish their own safety reports. Those reviews can be technically strong, but the lab normally controls model access, test timing and the scope of disclosure. A commercial evaluator can lose access or a contract. A public institution has a different job: set durable measurement practices, preserve institutional knowledge and explain high-consequence findings even when they are inconvenient.
OpenAI has itself called for strengthening CAISI as the federal government’s primary institution for frontier-model safety in a June policy blueprint. Support from a major lab is useful, but it also underscores why the center needs a defined mandate and publication policy that do not depend on any one company’s cooperation.
The next useful signals are concrete. CAISSI needs a permanent director, a public account of which evaluation categories continue, and a schedule for releasing methods and redacted findings where classified details cannot be shared. The administration also needs to define the division of labor between CAISSI, the Super Intelligence Force and the intelligence-led cyber benchmark.
Until then, the center is neither fully shuttered nor operating with the transparency its mission requires. Its recent cyber work shows that the technical capacity remains. The unresolved issue is whether that capacity will produce an independent public record, or mostly feed decisions made behind the White House and national-security boundary.
