CAISSI Staff Asked to Report to White House, AP Says | Zeniteq
Loading
Loading article
Zeniteq is now available on mobile for iPhone, iPad and Apple Watch.Download nowLearn more
CAISSI Staff Asked to Report to White House, AP Says
AP’s reporting raises questions about who directs U.S. AI safety evaluations, what findings become public, and how frontier systems will be independently tested.
The Center for Advancing Innovation and Standards for Super Intelligence faces significant operational instability, marked by rapid leadership turnover and a halt to public research. This lack of transparency complicates independent oversight of frontier AI models and autonomous agents, which require rigorous testing of intermediate behavioral steps rather than just final task outcomes. With only a ten-million-dollar budget, the center struggles to match the capacity needed for complex evaluations. These issues raise concerns that closer, informal White House oversight could politicize technical findings, potentially softening assessments of cybersecurity and biological risks to favor industry interests.
The U.S. government center responsible for independently assessing advanced AI systems has no permanent leader, has stopped publishing its safety research, and recently asked staff to report informally to a White House office, according to an Associated Press report published November 8, 2026.
The Center for Advancing Innovation and Standards for Super Intelligence, or CAISSI, sits within the Commerce Department. Independent evaluations can give policymakers and the public evidence beyond AI developers’ own descriptions of their systems.
AP’s reporting identifies three overlapping problems: unstable leadership, a missing public research record, and uncertainty about the center’s reporting relationship. Together, they make it harder to judge whether U.S. oversight is keeping pace with frontier models and increasingly autonomous AI agents.
What AP Reported About CAISSI
Two former government officials with knowledge of ongoing conversations told AP that CAISSI officials were recently asked to report informally to a White House office. They spoke anonymously to avoid repercussions for current staff.
The report identifies no public directive or specific White House office. The request’s scope, duration and place in the formal chain of command remain unclear. It should therefore be understood as a reported informal arrangement, not an announced transfer of the center out of Commerce.
A director was hired and fired within days in April.
His replacement served for a few months before resigning.
The head of CAISSI’s parent institute now also serves as its acting director.
Congress began asking in August why the center had stopped publicly releasing reports about AI systems’ safety performance.
Having an acting director means the center is not leaderless. The concern is continuity: repeated changes make it harder for outsiders to identify who owns the research agenda and is accountable for publication decisions.
The publication halt also needs a precise reading. AP reports that public research has stopped; that is different from evidence that all internal testing has stopped. The immediate, observable loss is transparency into what the center is finding.
CAISSI’s position involves two different kinds of independence.
The first is independence from AI developers. A government evaluator can examine a system without relying entirely on the company’s choice of tests, interpretation of results or description of acceptable risk.
The second is the ability to preserve technical findings within government, including findings that may be inconvenient for an administration’s policy goals. AP describes CAISSI as an independent division of Commerce, which places it inside the executive branch. Its relationship with political leadership therefore matters even without a formal organizational change.
Critics interviewed by AP worry that closer White House involvement could politicize decisions involving cybersecurity and biological risks, or lead to findings being softened or withheld in deference to the AI industry. These are concerns about possible outcomes, not documented instances of findings being altered.
AP also reports the opposing argument: closer access to the president could give the center greater prominence and potentially more funding.
Both possibilities make the operating rules important. White House coordination could help an evaluator secure resources or get urgent findings in front of decision-makers. It could also create pressure over what gets tested or published. The useful distinction is whether policy officials coordinate around technical findings or gain influence over the findings themselves.
My view is that publication practices offer the clearest near-term accountability measure. An organizational chart cannot reveal whether an unfavorable result survives review. A consistent public record of methods, limitations and findings gives outsiders something they can actually examine.
A Small Budget Adds Capacity Questions
AP reports that CAISSI receives $10 million annually, approximately one-eighth of the annual budget of its U.K. counterpart. That comparison gives a sense of scale, although budgets alone cannot establish whether two organizations have equivalent responsibilities or testing capacity.
Independent evaluation requires more than obtaining a model and asking it difficult questions. Evaluators need access, suitable test environments, technical expertise and time to investigate failures. Agent testing can add further work because researchers must examine sequences of actions and their consequences.
The practical question is how much coverage the center can provide: which systems it evaluates, which risks it investigates, and whether it can revisit findings as models and deployment conditions change.
AP reports that OpenAI had also called for greater funding and support for the division. That is relevant industry support for government evaluation, but it is not evidence that the center’s resources are sufficient.
The stronger conclusion from the published evidence is narrower: CAISSI has a modest allocation compared with its U.K. counterpart, while its leadership and public reporting are unsettled. The available reporting does not provide a workload breakdown that would support a precise estimate of the funding shortfall.
AI Agents Expose a Testing Gap
AI agents can plan and carry out multiple steps using tools and services. Evaluating them therefore requires attention to their behavior throughout a task, as well as the final result.
Sayash Kapoor, an incoming assistant professor at the University of California, Berkeley School of Information, told AP that there is still no standard process for testing how safely agents act in real-world situations, even for something as ordinary as booking a flight.
A completed booking can look like success in a final-result check. It does not, by itself, describe every action the agent took to get there. Kapoor emphasized examining “the entire set of steps that an agent took.” His comments appear in the AP report carried by ABC News.
That distinction has practical consequences. Consider an illustrative booking evaluation: an assessor might check whether the agent stayed within its authorized spending limit, used only permitted information and sought confirmation before making an irreversible purchase. Those are examples of evaluation questions, not findings about a particular product.
A useful safety report would explain:
The configuration: which model, tools and permissions were tested.
The conditions: what instructions, constraints and opportunities for human intervention applied.
The behavior: what actions occurred during the task.
The limits: what the evaluation could not show.
This is also why a capability rating and an independent safety assessment answer different questions. Coverage such as OpenAI’s GPT-6 Cyber critical cyber rating concerns consequential AI capabilities; oversight also needs to establish how systems behave under defined access and deployment conditions.
The analytical connection to CAISSI is straightforward: when agent evaluations require more detailed evidence, losing public research makes external scrutiny harder. Readers cannot compare methods or assess coverage if those details remain unavailable.
The White House Response Leaves Testing Unresolved
The White House did not answer AP’s specific questions about CAISSI. It instead pointed to President Donald Trump’s newly established Super Intelligence Force and a risk report commissioned from the President’s Council of Advisors on Science and Technology. Its statement emphasized continued U.S. leadership in the technology.
Those initiatives describe coordination and risk assessment. The response, as reported, did not explain who controls CAISSI’s research agenda, why publication stopped or when public reports might resume.
AP also reports that Trump asked leading AI companies in September to undertake “tremendous self-policing” through a voluntary safety pact. The administration had not explained how it would monitor implementation.
Voluntary cooperation can give government evaluators access to systems and information. Its value depends on what access companies provide, what evaluators can investigate and what happens when testing reveals a problem.
The next useful evidence would be operational: a clear reporting relationship, stable leadership, published evaluation methods and a defined process for communicating significant findings. For pre-deployment testing, timing matters too. An assessment completed after broad release serves a different purpose from one that can inform deployment decisions.
AP’s report leaves the central oversight question open: who will independently test frontier models and agents, under what conditions, and how much of the resulting evidence will the public be able to see?
Frequently Asked Questions
3 questions
1
Is CAISSI Still Part of the Commerce Department?
AP describes CAISSI as a division of Commerce. The reported informal White House arrangement is not an announced transfer to another department or office.
2
Does CAISSI Have Anyone in Charge?
Yes. AP reports that the head of its parent institute also serves as acting director. The vacancy concerns permanent leadership.
3
Does a Successful Agent Task Prove Safe Behavior?
No. A correct final result can leave important actions unexamined. Evaluators also need to assess the agent’s intermediate steps, permissions and compliance with task constraints.
