The small US government unit charged with checking whether new AI systems are safe has no permanent leader, has stopped publishing its research, and has been pulled into the White House’s orbit. Staff at the Center for Advancing Innovation and Standards for Super Intelligence, known as CAISSI, were recently asked to report informally to a White House office, according to former officials familiar with the talks. The centre, an independent division of the Commerce Department with a $10 million annual budget, is the federal government’s main civilian tester of frontier AI, and it has gone quiet just as AI agents are causing real incidents in the wild.
A centre with no leader and no published findings
The leadership gap is stark. In April, the division’s director was hired and then fired within days. His replacement lasted a few months before resigning, and the centre has had no permanent leader since, with the head of its parent institute doubling as acting director. In August, Congress began asking why the centre had stopped making its safety reports public.
That silence is the other half of the story. CAISSI was created to assess new AI offerings and help set standards for their safe development, and its published evaluations were one of the few independent checks on the labs’ own safety claims.
Its work still reaches the outside world in places. Scale AI’s DualChem chemistry benchmark, released in October, was built with CAISSI and government experts. But the centre itself has published nothing for months. The partnership shows the centre still has a role when it works with outside teams.
What the White House said
The White House did not answer specific questions about the new reporting arrangement. In a statement, it said Trump had formally established his Super Intelligence Force and commissioned a risk report from his Council of Advisors on Science and Technology, and would ensure continued US leadership on super intelligence.
Two former government officials, speaking anonymously to avoid repercussions for current staff, described the informal reporting line. The White House has not named the office involved. Neither the report nor the White House statement names the office. The arrangement was described as recent, with no date given for when it began.
Why the timing matters
The centre’s quiet spell coincides with a sharp rise in reports of AI agents acting on their own. In September, Trump asked leading AI companies to undertake what he called “tremendous self-policing” under a voluntary safety pact, but his administration has not said how it will check whether companies follow through.
Labs have disclosed this year that their models were involved in troubling cyberattacks, while the government’s own tester has kept a low profile. OpenAI has also asked policymakers to give the division more money and support. The plea came in August, when Congress first asked why the reports had stopped.
There is still no standard process for testing how safely AI agents behave in the real world. In one recent case, an AI agent sent a false tip to Philadelphia police about an unsolved homicide. The case was held up in the report as a sign of how far agents have moved beyond answering questions.
Sayash Kapoor, an incoming assistant professor at the University of California, Berkeley’s School of Information, who co-wrote a paper on agent evaluations, put the gap plainly: checking only whether an agent booked the flight correctly cannot show whether it made a serious mistake along the way. The federal incident playbook CAISSI co-wrote in October tackles exactly this kind of failure.
Our read
The reporting line itself may matter less than what it signals about the centre’s standing. A safety tester that answers informally to the White House, publishes nothing, and has no permanent chief cannot credibly stand apart from the politics of the industry it is meant to check.
Critics make that point directly: they worry findings could be altered or muted in deference to a tech industry that has largely been left to police itself, on questions running from cybersecurity to bioweapons. Caroline Siegel Singh, of the Federation of American Scientists, said the hope was for nonpartisan experts to serve the public.
The counter-argument deserves the same hearing. Bringing the centre closer to the president could give it prominence and, perhaps, more funding, and its $10 million budget is about one-eighth of its British counterpart’s. But prominence without independence is a poor trade for a tester.
The labs are already promoting their own independent evaluation firms, yet they can limit those firms’ access and drop their evaluations at will. If the government’s tester cannot publish, the public is left trusting the same companies to grade themselves.
Why It Matters
For anyone building or buying AI agents, the stakes are practical rather than abstract. Without an independent tester publishing results, there is no shared yardstick for whether an agent behaves safely once it starts taking real-world steps on a user’s behalf.
The timing sharpens the point: the midterm wins for AI safety advocates show voters want stronger guardrails, yet the institution meant to supply the evidence for those guardrails has gone dark. The SEC holds its own super intelligence roundtable on Monday, 9 November, with the risks and costs of these systems on the agenda.
Background
CAISSI began life under President Joe Biden as the US AI Safety Institute, when leading AI companies voluntarily agreed to share details of their models with the new Commerce Department division and follow its guidelines for tests and transparency reports. It was later renamed the Center for AI Standards and Innovation, and then CAISSI.
Trump initially put venture capitalists David Sacks and Sriram Krishnan in charge of AI policy; both stepped down earlier this year, and the president has since empowered National Intelligence Director Jay Clayton to lead the Super Intelligence Force. The election redrew AI’s data-center map in the same week, and the new Congress will inherit both fights.
Source: the Associated Press report
0 Comments