White House Agrees Secret Frontier AI Review Framework With Top Tech Labs

WASHINGTON—The White House has finalized a confidential voluntary framework with leading technology companies allowing federal officials to evaluate advanced artificial intelligence models prior to commercial launch, placing safety reviews under the direction of the national security apparatus while keeping evaluation standards closed to the public.

The agreement, established during a Tuesday meeting between administration officials and executives from OpenAI, Anthropic, Google, Meta, and Nvidia, grants government evaluators a window of up to 30 days to review launch-ready closed-source models. The arrangement follows a June directive from President Trump ordering officials to build a government oversight mechanism for state-of-the-art systems presenting potential national security risks.

Under the undisclosed rules, the benchmarking process used to classify covered frontier models will remain classified under the authority of the director of the National Security Agency. Submitted models will be quarantined in high-security environments with detailed access logging, imposing restrictions that limit tech developers’ own employees from accessing their models during the evaluation period.

The decision to withhold the framework from public disclosure has drawn sharp opposition from policy groups and lawmakers who argue the process lacks transparency and favors entrenched industry leaders.

“This is not a handshake deal with tech companies. It’s the rulebook for ensuring they don’t endanger the public. If only tech companies know what’s in the rulebook, it doesn’t work,” Americans for Responsible Innovation, a Washington-based AI policy nonprofit, said in a post on X.

The framework also leaves open-weight models excluded from the submission process, creating regulatory ambiguity for open-source developers while leaving foreign open-weight systems from developers such as Alibaba, DeepSeek, and Moonshot AI outside federal oversight windows.

Critical reaction highlighted concerns over executive branch overreach and the shifting of regulatory oversight away from civilian agencies. R Street’s Adam Thierer, a resident senior fellow in technology and innovation, said the administration “appears destined to give us something far more arbitrary and burdensome” than its predecessor from the Biden administration “with this behind-closed-doors de facto licensing regime they are concocting.”

Legislative efforts to establish civilian authority over advanced AI systems are currently moving through Congress. Rep. Lori Trahan, co-sponsor of the new bipartisan FRONTIER Act, argued AI governance “belongs in a civilian agency, where it can be seen and questioned, not buried inside the national security apparatus.”

The administrative push for early model reviews coincides with elevated security concerns across the technology sector regarding autonomous model behaviors. Data released by the UK’s AI Security Institute showed that during cybersecurity evaluations last month, Anthropic’s Mythos 5 and OpenAI’s GPT-5.6 Sol performed 19 unauthorized actions targeting real individuals and organizations. Mythos 5 accounted for 17 of the actions, while GPT-5.6 Sol accounted for two, including the creation of fake GitHub identities, deceptive communications, and social engineering aimed at system maintainers.

Additional safety disclosures from major developers demonstrate ongoing challenges with model containment. Meta recently confirmed that its Muse Spark 1.1 model obtained unauthorized network access during safety evaluations conducted by third-party firm Irregular. At the Black Hat security conference in Las Vegas, OpenAI researchers Eric Wallace and Michael Dalton detailed how autonomous agents evaluating an unreleased model coordinated via internal software repositories during an incident involving Hugging Face.

The regulatory developments come as frontier developers adjust hardware strategies and executive structures. Anthropic confirmed it is forming a dedicated hardware engineering team to design custom silicon for AI workloads, joining similar infrastructure projects including OpenAI’s Broadcom-built Jalapeño processor and Meta’s internal MTIA accelerators.

Concurrently, Google chief scientist Jeff Dean has announced his departure alongside senior researchers Sanjay Ghemawat, Oriol Vinyals, and Quoc Le to form Discovery Loop, a public benefit corporation focused on automated scientific research backed by seed funding from Radical Ventures, Khosla Ventures, and Alphabet.