Safety talks need no waiver, says OpenAI's Lehane

Dario Amodei spent part of his weekend essay asking Washington for a narrow antitrust waiver, so that rival labs could coordinate on safety without legal trouble. On Tuesday OpenAI's global policy chief Chris Lehane said the labs do not need one, and, as TechCrunch reports, that OpenAI, Anthropic and Google DeepMind have already been talking for weeks.

At a glance

• Chris Lehane, OpenAI's global policy chief, told reporters on Tuesday that the company has spent weeks working with Anthropic and Google DeepMind on AI safety, confirming coordination that previously had only been hinted at.
• Lehane said OpenAI supports a provision in the FRONTIER Act that would force top frontier labs to let "independent verification organizations" into their companies to check that models are developed safely.
• Amodei's essay proposed a narrow government waiver to keep safety coordination clear of antitrust law; Lehane reportedly said the firms do not need one, and Altman has flagged the same legal risk.

If you were offline this weekend: on Saturday Amodei published an essay urging the industry to work together to slow the pace of frontier AI and avoid catastrophic risks. According to Shattered, the essay is titled "We Must Pace the Frontier," runs roughly 3,800 words, and is explicit that pacing does not mean halting model training or technical progress. AI leaders including Sam Altman, Demis Hassabis and Elon Musk backed the call.

Amodei asked for a waiver; Lehane says none is needed

Lehane told reporters on Tuesday, in remarks first reported by Bloomberg, that OpenAI has been working with Anthropic and Google DeepMind on safety for weeks. He is reportedly in Washington to work with U.S. lawmakers on the catastrophic risks associated with AI.

The essay proposed a narrow government waiver for exactly this kind of coordination. Lehane reportedly said the firms do not need one, even though Altman has noted such talks could put the companies at risk of violating antitrust law if the coordination is found to suppress competition. In a Fortune interview late last week, Altman hinted he had been in private discussions with other AI leaders.

According to the essay on his personal site, Amodei put the demand in one line: "We must slow the pace at which we improve the capabilities of AI models," adding that progress will still seem fast and that the industry must make wise use of the time it gains.

The Information reports the three labs are building a standards body

This week The Information reported that OpenAI, Anthropic and Google DeepMind have been working together to create a standards body for the AI industry. Altman reportedly told staff that this would have to happen without the support of the U.S. government.

Hassabis called on the U.S. in July to set up a new standards body to act as an AI watchdog, with the power to screen the world's most advanced models and to coordinate industrywide slowdowns if the dangers progressed.

President Trump has dismissed safety concerns as a hoax and pushed back on tighter regulation, saying any slowdown would give China a leg up in the race. His key AI advisor David Sacks, a longtime investor with a personal stake in much of the industry, has said fears of existential risk from AI are overblown.

OpenAI backs a FRONTIER Act provision on outside verifiers

At the same Tuesday meeting, Lehane said OpenAI supports a provision in the FRONTIER Act that would force top frontier labs to allow "independent verification organizations" into their companies to ensure models are developed safely.

That tracks with what Altman said when he backed Amodei's call: OpenAI would join Anthropic in embedding third-party evaluators inside the company to monitor for safety. Hassabis and Musk also supported the essay and its call to action.

What does an outside verifier do once it is inside a lab?

The Frontier Model Forum's own material describes three functions for third-party assessment: Confirmation, checking whether a developer's safety assessments are accurate and well-supported; Robustness, applying independent methods including adversarial testing to elicit harmful behaviour; and Supplementation, adding specialist expertise the lab does not have in-house.

Per the same Frontier Model Forum material, internal work runs in two stages, Frontier Capability Assessments and Mitigations Assessments, and outside assessors come in at times and depths calibrated to the model, the deployment context and the risk. In plain words, the lab tests first, and the outsider re-checks those conclusions, tries to break the model itself, and covers the ground the lab cannot.

None of this is an agreement yet. The reporting gives no signed text, no membership or launch date for the standards body, and no reasoning behind Lehane's conclusion that no waiver is needed. In our view the gap between Amodei asking for a government carve-out and OpenAI saying none is required is the detail to watch, since it decides whether the talks stay informal or turn into a written regime.

Where the FRONTIER Act lands

The checkpoints here are legislative and unscheduled. Whether the verification provision Lehane endorsed survives in the FRONTIER Act, whether any lab ultimately asks Washington for the waiver Amodei proposed, and whether the standards body The Information described gets a name, a charter or members are all still open. No date has been given for any of that.