Vetted AI labs can see their own classified safety scores. The rest of the field ships blind.

Federal AI safety testing is becoming a private certification track. The federal government runs the benchmark, the benchmark is classified, and only the developers who can see their own results get to act on them. Everyone else ships into the dark.

WIRED's reporting on the new White House framework names the mechanism. A 30-day voluntary pre-release submission window feeds a classified benchmarking step, and the vetted scores flow to federal agencies alongside a closed list of "trusted corporate partners." Open-weight models are reportedly excluded. Smaller labs are not on the list. Outside researchers cannot see the test, cannot submit into it, and do not get the results.

Read against the administration's stated rationale, that the framework is narrow, that it touches only the most capable frontier systems, and that secrecy is justified by national security, the structural picture reads as entrenchment by structure. The government does not need to pick winners. The information asymmetry does it. The lab that knows its score can fix its model, brief its customers, and price the gap. The lab that does not is shipping blind.

Brad Carson, who leads a nonprofit, told WIRED the secrecy itself is the issue. The narrower counter, that the framework only touches frontier systems, is the more durable concern. When safety testing becomes a voluntary submission and trusted-partner results lane, it creates structural information asymmetry.