WASHINGTON – The White House is nearing a voluntary agreement with OpenAI, Anthropic, and Google that would govern how the next generation of AI models gets tested before public release. The standards were drafted in close consultation with the very labs they are meant to hold accountable.
The talks, first reported by the Financial Times on July 3, represent the furthest the federal government has come toward formalizing its oversight of the AI industry’s most powerful systems. An announcement could come as soon as the first week of August, when a 60-day deadline set by President Trump’s June 2 executive order on AI and cybersecurity expires. The order directed Treasury, Defense, and Homeland Security to develop a benchmarking process and voluntary framework for advanced AI models by that date.
What the framework is supposed to do reads simply enough on paper. AI developers would give the federal government access to their newest models up to 30 days before public release. The Center for AI Standards and Innovation, housed within the Commerce Department, and the National Security Agency would run classified benchmarks assessing each model’s advanced cyber capabilities. Conditions on who gets to use the models, among them domestic commercial users, vetted foreign companies, and foreign governments, would be part of the access structure attached to each release. Commerce Secretary Howard Lutnick has already been directly involved in at least one launch: AI Weekly reported he personally approved roughly 20 vetted customers during a limited preview of GPT-5.6 while OpenAI held back a full public release at the administration’s request.
That last detail matters because the framework is not arriving in a vacuum. The government has already intervened in real model launches before any formal rulebook existed.
Anthropic got the sharpest demonstration of what informal government leverage looks like in this space. On June 12, the Pentagon issued an export control order targeting Anthropic’s most capable models after a South Korean telecommunications company with Chinese ownership ties attempted to access them. The Claude systems were effectively locked for foreign commercial customers for nearly three weeks. The controls were lifted July 1, on the condition that Anthropic agreed to implement what the company described as “99%-plus jailbreak filters.” That condition did not come from any published standard. It was set by the government, accepted by the company in private, and has not been documented in any form that would allow independent verification of what 99% means or how it would be measured.
The talks also involve Amazon, Microsoft, and several other labs beyond the three principal parties the Financial Times named. Notably absent is Meta, which Gizmodo reported has been resistant to or excluded from the framework. That gap is significant: Meta’s Llama model family is open-weight, meaning the model weights are publicly distributed and cannot be meaningfully constrained by any access control applied at the lab level. A framework covering the three walled-garden labs while the most widely used open-weight models remain entirely outside it has a structural hole in it from day one.
The framework also has to reckon with what “voluntary” actually means here. The executive order explicitly prohibits the government from establishing a “mandatory governmental licensing, preclearance, or permitting requirement” for AI model development or release. That language was apparently included to reassure the industry that Washington was not building a regulatory approval process. It also means the enforcement mechanism behind the framework consists of the same informal pressure the administration already demonstrated on OpenAI and Anthropic: export control threats, delayed launch approvals, and direct calls from cabinet officials.
Researchers at the Accountable AI Lab at the Wharton School of the University of Pennsylvania described the arrangement as “a backdoor licensing regime built on existing Commerce Department authority, with conditions that shift without public notice.” That framing points at something specific: the benchmarks the government uses to evaluate frontier models are classified. Labs will not know precisely what their systems are being tested against. Independent researchers cannot audit the results. And the conditions for clearance can be set and revised in private, as Anthropic’s June experience made clear, with no public record of what changed or why.
The financial stakes running underneath these negotiations are substantial. As Eastern Herald reported last week, OpenAI separately pitched the administration a 5 percent equity stake worth $42.6 billion as a mechanism for reducing political scrutiny of the AI industry. That conversation is running parallel to the frontier-model standards talks and involves overlapping officials. What the administration is trading away in each negotiation, and what it is receiving in return, remains outside public view.
Meanwhile, Anthropic is exploring hardware independence through talks with Samsung over a custom chip at the 2nm process node. The labs now negotiating these standards are simultaneously deepening infrastructure investments that would make any rupture with Washington substantially more costly. That alignment of financial exposure with regulatory compliance is not accidental.
There is something worth sitting with in the structure of what is being finalized. The classified benchmarks that will determine whether a model gets cleared for release are being developed by agencies that report to the executive branch, using criteria the public will not see, applied to models submitted by companies that helped draft other provisions of the same framework. When the administration announces these standards before August 1, the most important question is one the announcement is unlikely to answer: if a model passes the government’s classified tests and causes serious harm after release, who is accountable, and against what standard were they supposed to be held?
