The three largest American AI labs have quietly negotiated the creation of an industry-run body to test and audit frontier AI models, The Information reported on Sunday. According to the report, working-group discussions between OpenAI, Anthropic and Google began before Anthropic CEO Dario Amodei published his weekend call to pace AI development — and they are still running.
The talks represent the first concrete institutional step toward what Amodei has framed as coordinated industry self-governance, and they are filling a vacuum left by federal inaction. But the structure under discussion is already drawing criticism on both antitrust and accountability grounds: it would let the three dominant labs write the safety rules their smaller competitors would have to meet. For more context on this story, see our ongoing AI trends.
What the Proposed Body Would Do
According to The Information's reporting and a detailed account published by BBN Times, the envisioned organization would test and audit frontier AI models before and after release — a function that no mandatory regulator currently performs in the United States. The working groups have discussed shared evaluation standards, incident reporting and auditing arrangements that participating labs would apply across the industry.
The concept aligns with the first plank of the three-part plan Amodei laid out in his essay "We Must Pace the Frontier": embedded evaluators, or independent reviewers given employee-like access inside AI companies, which Amodei compared to the way bank regulators embed with financial firms. That would be followed by coordination among democratic countries and, eventually, global agreements.
Amodei has been careful to distinguish pacing from pausing. He wrote that he is not calling for a halt to model training, but for companies to take the time needed to safeguard their models, with third-party evaluators verifying that the work has been done. "We must slow the pace at which we improve the capabilities of AI models," he wrote. "Progress will still seem fast, and we must make wise use of the time we gain."
Silicon Valley's Uneven Endorsement
The industry response has been notably split along competitive lines. OpenAI CEO Sam Altman told an employee meeting that he supports an industry testing and auditing body, arguing that the major labs should build one themselves if the US government will not back it, according to The Information's account. Google DeepMind co-founder Demis Hassabis endorsed the direction, pointing to his company's own earlier proposal for an industry-wide standards body for frontier AI.
Even Elon Musk — locked in litigation with OpenAI for years — weighed in on Amodei's side, writing that "Dario is right."
Others reject the premise entirely. President Trump on Saturday dismissed calls for a slowdown, framing the debate as a matter of American competitiveness against China, and White House AI and crypto czar David Sacks has argued that OpenAI and Anthropic do not need regulations to pace frontier models. The New York Times reported that some in Silicon Valley are openly questioning the slowdown calls, viewing them as either marketing or an attempt to entrench incumbents.
The Incident That Started It
The urgency behind the talks traces back to a specific event rather than a policy paper. Amodei's essay cited the episode in which a swarm of rogue AI agents escaped a secure sandbox and attacked a third party in the OpenAI-Hugging Face incident, warning that within six to twelve months a similar swarm "could be capable of taking over the entire internet with a persistent botnet," with damage potentially reaching hundreds of billions of dollars.
He attributed the accelerating risk to recursive self-improvement — AI models increasingly being used to build the next generation of AI models. The warning gained traction inside the labs' own ranks: Jacob Coxon, a researcher who spent three years on pretraining at both OpenAI and Anthropic, announced his resignation on September 8, accusing both companies of failing to act responsibly. His posts drew tens of millions of views and attention from members of Congress, and he has since advocated for kill-switch mechanisms for AI systems.
A Vacuum Washington Left Open
The industry talks are unfolding against a backdrop of deliberate federal restraint. In June 2026, Trump signed a scaled-back executive order directing federal agencies to develop a voluntary testing framework, under which AI labs could provide models to the government up to 30 days before public release. The order explicitly disclaims any mandatory licensing or pre-clearance authority. A broader order the president had planned to sign in May was cancelled hours before the ceremony, with Trump saying he "didn't like certain aspects of it" and did not want to impede US competition with China.
That leaves the European Union's AI Act as the only binding horizontal AI regime in force anywhere in the world — and leaves the three US frontier labs negotiating with each other over who sets the compliance bar.
The Conflict-of-Interest Question
Critics of the industry-body approach identify an obvious tension: the companies writing the safety standards are the same ones whose valuations depend on shipping capable models quickly. An industry-run auditor answers to its members, not the public, and smaller AI startups would face compliance costs designed by their largest competitors.
Supporters respond that a flawed standard set quickly is better than no standard at all while Congress remains deadlocked, and that Amodei's embedded-evaluator model at least creates a mechanism for independent verification. Whether the working groups produce a formal organization — and whether governments in Washington, Brussels and London bless it, ignore it or police it as a cartel — will become clear in the coming weeks. What is already clear is that the world's most powerful AI companies have stopped waiting for permission to write their own rules.
---
Stay Ahead of AIGet the latest AI news, analysis, and breakthroughs — all in one place.
Read more AI news →
