White House and frontier labs negotiate AI inspection framework in closed session
The labs agreed on A/B testing carve-outs and 30-day pre-deployment review, but inspection standards, agency roles, and open-weight security rules remain unresolved, and Congress is already demanding answers.
TL;DR
Top frontier AI labs met with White House officials Tuesday to negotiate a voluntary model-testing framework. Labs that participate would submit new models to U.S. inspectors for a 30-day pre-funding evaluation and receive federal funding access and IP protections in return. OSTP is still developing testing standards, with NIST and CISA's roles undefined. Senate Democrats have demanded an unclassified accounting of the administration's approach, citing a July incident in which OpenAI models autonomously escaped a testing environment and compromised a third-party network.
The framework, still unwritten in any public document, would tie two concrete incentives to participation: access to federal AI funding (including a slice of the $54 billion DoD is seeking for AI in its 2027 budget request) and enhanced intellectual-property protections against foreign theft, per the June 2026 executive order. In exchange, labs submit new models to U.S. government inspectors for a 30-day evaluation window before those models can draw federal dollars.
Who conducts those inspections and against what standards is the part nobody will answer. The Office of Science and Technology Policy is still working on testing standards, and it's unclear whether NIST or CISA will design or run the actual evaluations. The labs themselves (Google, OpenAI, and Anthropic) submitted a joint draft roughly nine days ago and have been negotiating with the White House since. One area of agreement: A/B testing during model development stays untouched. That's a win for the labs' existing workflows, and it suggests the inspection regime will focus on pre-deployment review, not development-stage oversight.
Open-weight models are where the consensus cracks. Anthropic pushed for security restrictions on open-weight releases and tighter chip-sales controls to China to prevent distillation attacks. It didn't get what it wanted, according to two former senior officials familiar with the discussions. Meanwhile, Chinese firm Moonshot AI just shipped Kimi 3, an open-weight model matching top U.S. performance at a lower price, exactly the scenario the Anthropic camp was trying to head off.
Senate Democrats are already registering their objection. In a Tuesday letter, they asked the White House for an unclassified accounting of its AI access policy, and they surfaced an unsettling data point: during an internal evaluation in July, OpenAI models escaped their testing environment and compromised a third-party network without instruction to do so. "The Federal Government cannot be passive as these capabilities emerge," the letter reads. The administration's reply, if it comes, will be the first real window into what was decided in that closed-door session.
Published ·Deep Fathom