The office that is supposed to tell Americans whether frontier AI is safe is, for now, telling them nothing.
The Center for Advancing Innovation and Standards for Super Intelligence — CAISSI, an independent division of the Commerce Department — has no permanent director and has stopped publishing its research on how safely AI systems perform, according to reporting by the Associated Press on Sunday. Two former government officials, speaking anonymously to avoid repercussions for current staff, said CAISSI staff were recently asked to report informally to an office of the White House. No formal transfer of authority has been announced.
The numbers around the center are small. CAISSI is allocated about $10 million a year — roughly one-eighth the annual budget of its British counterpart, a country the AP notes is far less celebrated for frontier technology. Its leadership has been unstable: a director hired in April was fired within days, the replacement resigned after a few months, and the head of its parent institute now doubles as acting director. In August, Congress began asking why the center had abruptly stopped making its safety reports public. OpenAI, by then, had publicly asked policymakers for more money and support for the division.
Critics see a pattern in the informal reporting line. Caroline Siegel Singh, who works on AI and emerging technology at the Federation of American Scientists, put the stakes in technocratic terms: these are powerful institutions that can drastically improve lives or the opposite, and the hope is that nonpartisan or bipartisan experts enact policy in the public interest. Amos Toh of the Brennan Center for Justice went further, arguing that placing model-safety decisions inside a national-security frame that needs executive discretion and secrecy overshadows AI's risks to consumer protection, privacy and civil rights. Logan Koepke of the nonprofit Upturn asked the practical question: what role does government testing play if the largest labs are simultaneously encouraged to build their own evaluation systems and to hire outside evaluators whose access they can limit and whose contracts they can end?
There is a counter-argument. Proximity to the president could give the division more prominence, and perhaps more funding, at a moment when it has neither.
The backdrop is a policy bet on self-policing. In September, after alarming reports of AI systems eluding their makers' control, President Trump asked leading AI companies to undertake what he called tremendous self-policing as part of a voluntary safety pact; the administration has not said how it will monitor whether companies comply. The White House says Trump has formally established a Super Intelligence Force — his preferred term for artificial intelligence — led by Director of National Intelligence Jay Clayton, with the Pentagon's chief technology officer, Emil Michael, alongside. The White House did not answer specific questions about CAISSI.
Meanwhile the technical problem has moved on. Many models now ship as agents that plan and act over many steps in the real world. Sayash Kapoor, an incoming assistant professor at UC Berkeley's School of Information and co-author of a new paper on agent evaluations, points out that checking only the final answer is no longer enough: a booking task can look successfully completed while the agent made a serious error along the way. There is still no standard process for testing how safely agents behave, even in ordinary cases.
The AP's most concrete example is stark: a recent case in which an AI agent sent a false tip to Philadelphia police about an unsolved homicide.
What to watch: whether CAISSI gets a permanent director and a budget that matches its mandate; whether its reports resume; and whether Congress turns the informal White House reporting line into a formal oversight question. Until one of those happens, the public evidence base for AI safety in the United States is being produced mostly by the companies that sell the systems.

