OpenAI, Anthropic and Google have been negotiating an AI safety standards body since July -- and Meta told the White House not to join

OpenAI's global policy chief, Chris Lehane, told reporters Monday that OpenAI, Anthropic and Google DeepMind have been coordinating on AI safety "for weeks" -- the first on-the-record confirmation that talks the three labs had not discussed in public trace back well before Dario Amodei's September call to slow down. Lehane added that the companies don't need a government waiver to keep talking, which quietly undercuts the frame of Amodei's own essay two days earlier, which asked democratic governments to help coordinate exactly this kind of discussion. The reporting lands three days after President Trump dismissed Amodei's warning as a "SICK conspiracy" on Truth Social -- the labs, it turns out, hadn't waited for Washington's answer either way.

The shape of what they're discussing predates Amodei's essay by two months. In July, Google DeepMind chief Demis Hassabis published a proposal for a US-led AI standards body modeled on FINRA, the private association that polices Wall Street brokers: a public-private partnership, government-overseen but industry-funded, staffed by "independent leading technical experts and open-source representatives" who would test frontier models before release for dangerous cyber, biological and deception capabilities. Treasury Secretary Scott Bessent was reportedly weighing a FINRA-like agency reporting to the SEC within days of the proposal -- the closest the idea has come to an actual government hook, and still short of one.

(FINRA doesn't test a brokerage's products before they go to market -- it examines member firms after the fact and fines the ones that broke the rules. An AI standards body modeled on it, testing models before release, would actually be a bigger ask than the FINRA comparison suggests.)

Amodei's own essay lays out three steps, and only the first is a firm commitment: Anthropic is unilaterally granting outside evaluators "employee-level access" to its systems, with publication rights the company says it won't edit. Step two -- common safety standards among labs in democratic countries -- is the part Lehane, Hassabis and now Bessent are all circling, and it is still voluntary. Step three, coordinating pacing with authoritarian governments including China's, is the part nobody has begun.

OpenAI chief scientist Jakub Pachocki went further than a polite nod: "I believe that shared safety standards and international coordination on further AI development need to be priorities now," he said, while Sam Altman said the company would "look forward to collaborating with our colleagues across the industry" on a shared standard. Neither has said what OpenAI would actually accept as binding, or what it would do if a future standard flagged one of its own models.

The one frontier lab conspicuously outside the conversation is Meta. Mark Zuckerberg is reported to have personally advised President Trump against the standards-body idea -- putting Meta on the same side of the argument as an administration that has dismissed Amodei's slowdown call as manufactured risk, even though Meta has never disputed that its own models could pose the risks Amodei describes.

The proposal has real critics, and their objection isn't abstract. Gartner analyst Nader Henein called industry self-regulation "not viable," arguing most AI vendors lack the capacity to police themselves without a conflict of interest baked in. Independent analyst Carmi Levy was blunter about Hassabis's framework specifically.

“A self-serving roadmap for an industry bent on racing to the AI horizon regardless of the harms caused along the way.” — Carmi Levy, independent technology analyst

Yoshua Bengio, the Turing Award-winning researcher who now runs the nonprofit LawZero, focused on the plan's biggest structural gap: it starts voluntary, with no stated path to becoming mandatory, and it wouldn't reach models built for military or government use at all.

The irony sits in a number none of the three labs raised on the record. The Future of Life Institute's Summer 2026 AI Safety Index graded nine labs on risk assessment, safety frameworks and governance using the standard US GPA scale -- reviewed through June 3, before Hassabis's proposal existed. No lab in the industry scored above a C+. Anthropic topped the field at 2.66; Google DeepMind and OpenAI, the two labs it's now negotiating a shared standard with, scored 2.01 and 2.28, both a C. Meta, the lab sitting the conversation out, scored 1.32 -- a D+.

None of this makes the standards-body idea worthless. It is the first time OpenAI, Anthropic and Google DeepMind have described the same effort in public, on the record, in the same week -- more coordination than the industry has managed on anything since the original White House safety commitments of 2023. What's missing is anyone saying, specifically, what happens to a lab that signs on and then ships a model the body flags.

Whether that gap closes is the actual test. A voluntary standard written and funded by the three companies with the most to lose from a binding one is not nothing -- but it isn't the thing Amodei's essay asked for, either. His own three-step plan puts common standards as step two of three, with step three -- pacing coordination with authoritarian governments -- still entirely theoretical. The industry has, for the first time, admitted step two is already underway. It has not yet said what makes it more than a promise the participants wrote for themselves.