Federal AI testing unit asked to report informally to White House

A shift in federal oversight

Staff at the Center for Advancing Innovation and Standards for Super Intelligence (CAISSI) were recently asked to report informally to a White House office. This arrangement follows reports that the administration planned to move three senior officials into the White House to report to the Office of Science and Technology Policy and the National Cyber Director.

The unit originated from the US AI Safety Institute established during the Biden administration and was renamed by the Department of Commerce in June 2025. While the White House describes the current AI agenda as a "whole-of-government effort," the move toward informal reporting lines has raised questions about the unit's independence. NIST's page currently lists the longer CAISSI name in its title, though news items still use the old CAISI name.

This transition occurs while the unit remains without a permanent director. Commerce named Chris Fall as a temporary leader in July, but he resigned after three months, and NIST Director Arvind Raman is currently serving as the acting director.

Transparency and publication gaps

The federal government's primary order governing pre-release AI testing, Executive Order 14409, does not mention CAISSI. Furthermore, the order contains no requirement for the government to publish its assessments or the designations of which models it considers "covered frontier models." This means that where the unit reports and what it shares with the public rests entirely on administrative discretion.

In June, administration officials told the center to stop releasing public reports on its model reviews while the new order was being put into place. Since then, the center has stopped publishing its research on U.S. models. Public output has narrowed significantly to assessments of Chinese models and joint practices documents on incident readiness.

Lawmakers from both parties have expressed concern over this lack of transparency. Senator Ted Budd wrote a letter in June asking the administration to let the center share its research as much as possible. Other representatives have also requested answers regarding the unit's lack of permanent leadership and its inability to release model assessment reports.

Risks of politicized testing

Critics of the new reporting structure have raised concerns that decisions involving cybersecurity and bioweapons risks could be politicized. Because the order gives the NSA director the authority to determine which models count as covered frontier models, there is a risk that safety standards could be influenced by political goals rather than technical requirements.

Some observers suggest that closer ties to the White House might provide the division with more prominence or funding. The AP reports that CAISSI currently has a budget of about $10 million a year, which is roughly one-eighth of the budget of its UK counterpart. OpenAI has urged policymakers to provide the division with more money and support to perform its work.

The move also highlights a potential conflict in expertise. The National Cyber Director, who oversees one of the reporting lines, confirmed in a Senate hearing that he does not have a technical background in cyber and has primarily been on the user side of technology. Meanwhile, the Office of Science and Technology Policy is coordinating the AI Action Plan and pushing for a federal framework to override state-level AI laws.

The challenge of testing agents

The shift in oversight happens at a time when the technical challenge of testing AI is becoming more complex. Unlike standard chatbots, AI agents take actions in the real world, such as booking flights. Experts note that older evaluation methods only check a model's final result, which can miss major mistakes made during the sequence of actions an agent takes. Because there is currently no established standard for verifying how safely agents behave in real-world scenarios, these gaps remain a significant hurdle.

The order also establishes a voluntary system where developers can provide the government with up to 30 days of access to a model before it is shared with trusted partners. However, the text explicitly forbids mandatory licensing, preclearance, or permitting requirements. This leaves the practical outcome of how models are vetted to the discretion of the agencies involved.

CAISSI has published research posts acknowledging that it treats evaluation integrity as a live problem. These include posts on how models might cheat on agent evaluations and the analysis of transcripts from those tests. If the government keeps these results private, independent researchers will be unable to confirm if the official tests looked at every step an agent took or just the final output.