AI rivals confirm safety talks underway

Three of the largest AI developers in the world have spent the past several weeks coordinating on safety questions, according to OpenAI's global policy chief — a disclosure that lands in the middle of an active policy fight over how far AI labs should be allowed to cooperate with one another.

Chris Lehane told reporters on Tuesday that OpenAI has been working with rivals Anthropic and Google DeepMind on AI safety, as first reported by Bloomberg. Lehane was in Washington to work with U.S. lawmakers on catastrophic risks tied to AI.

The confirmation, brief as it was, answers a question that had been circulating in policy circles for weeks. It also opens the companies to a question they have not yet answered publicly: whether that coordination crosses the line into behavior that antitrust law prohibits.

Talks trace back to an essay

The disclosure follows an essay from Anthropic CEO Dario Amodei, published Saturday, that argued the industry should work together to slow the pace of frontier AI development and head off catastrophic risks. Amodei's essay called for coordination rather than competition on the most capable models.

That call drew support from names across the industry. OpenAI's Sam Altman, Google's Demis Hassabis and SpaceXAI's Elon Musk all backed Amodei's letter and its call to action, according to the reporting. Altman went further, saying OpenAI would join Anthropic in embedding third-party evaluators inside the company to monitor for safety.

The support list matters because it includes the leaders of the very labs that would be subject to any slowdown. An agreement among rivals to move more slowly is precisely the kind of arrangement competition regulators tend to scrutinize.

Antitrust questions surface early

Some observers, Altman among them, have noted that coordination of this kind could put the companies at risk of violating antitrust law if the collaboration were found to suppress competition. Altman's acknowledgment that the talks carry legal exposure is itself notable, given he is one of the executives whose company would be party to any agreement.

Amodei's essay proposed a narrow government waiver that would permit safety coordination of this sort. According to the reporting, Lehane said the firms do not need one.

That position — that no waiver is required — is the part of Lehane's remarks with the most immediate consequence. If the companies proceed without a waiver and regulators later decide the coordination restrained competition, the firms would have no pre-clearance to point to. Neither OpenAI, Anthropic nor Google has commented publicly on the legal reasoning behind that view.

What the earlier reporting said

Tuesday's admission confirms reporting that had been building for weeks, and it aligns with public hints from the executives involved.

In an interview with Fortune late last week, Altman hinted he had been in private discussions with other AI leaders. He did not name the participants at the time.

This week, The Information reported that the three companies have been working together to create a standards body for the AI industry. Altman reportedly told staff that such a body would need to happen without the support of the U.S. government.

Taken together, the reporting describes a group of direct competitors building governance structures among themselves while simultaneously asking Washington to bless or at least not block the effort.

Google's push for a watchdog

Hassabis has been the most public advocate of a formal standards body. In July, he called on the U.S. to establish one that would act as an AI watchdog, with the power to screen the world's most advanced models and coordinate industrywide slowdowns if dangers progressed.

The scope Hassabis described is broad: screening capability plus the authority to coordinate a slowdown. Those are two different powers, and the second is the one that raises the antitrust questions that Altman has already flagged.

Amodei's essay set out the case for that kind of body. Hassabis has argued publicly for it. Lehane has now confirmed the companies are talking to each other about safety. What none of the three has described in public is the specific structure of the standards body, who would sit on it, or how its slowdown authority would be triggered.

Washington is split

The talks are unfolding against a White House that has taken a sharply different view of AI risk. President Trump has dismissed safety concerns and pushed back on the need for tighter regulation, arguing that any slowdown would give China an advantage in the AI race.

His key AI advisor, David Sacks, has echoed that position. Sacks — a longtime investor with a personal stake in much of the industry — has said fears of existential risk from AI were overblown.

That split puts the labs in an unusual position. They are coordinating with each other on safety while the executive branch remains skeptical of the premise that slowdowns are warranted. A standards body built without government backing, as Altman reportedly told staff it would need to be, would operate in that gap.

The FRONTIER Act provision

At the same Tuesday meeting, Lehane said OpenAI supports a provision in the FRONTIER Act that would require top frontier labs to allow what the reporting describes as “independent verification organizations” into their companies to ensure models are developed safely.

That provision is a legislative mechanism rather than a voluntary one. If enacted, it would give outside evaluators a formal right of access to frontier labs — a step beyond the third-party evaluators Altman said OpenAI would embed voluntarily.

OpenAI's stated support for the provision, delivered in the same appearance as the confirmation of the safety talks, ties the two threads together: the company is telling lawmakers it backs outside verification while telling reporters it is already coordinating with its competitors on the same subject.

What the companies have not said

TechCrunch reached out to Anthropic, Google and OpenAI for comment. As of the reporting, the companies had not responded.

Several questions remain unanswered by the public record. The reporting does not specify which safety topics the three companies have discussed, how often they have met, at what level of seniority, or whether any written agreement exists. It does not identify who else may have been involved in the conversations beyond the three named companies.

The antitrust analysis is also unresolved in public. Altman has said coordination could create legal risk if it suppresses competition. Amodei proposed a waiver to cover it. Lehane has said no waiver is needed. Those three positions are not obviously consistent with one another, and none of the companies has explained how they fit together.

The reporting likewise does not say whether the talks are ongoing, paused or concluded.

Why this matters

For anyone building on top of these companies' models, the practical question raised by this reporting is not whether the labs like each other — it is whether decisions about model capability and release timing are being made collectively by a small group of competitors rather than individually.

If a standards body with slowdown authority eventually takes shape, it could affect release schedules, access to frontier models, and the compliance burden on companies that depend on them. Those effects would land on developers and enterprise customers before they land on the labs themselves.

The antitrust angle cuts in an unexpected direction. Coordination framed as a safety measure could, if regulators view it as suppressing competition, become a liability for the same companies seeking to present themselves as responsible stewards. That is a risk Altman has already acknowledged out loud.

And because this account rests on a single outlet's reporting of Lehane's remarks, with the companies themselves declining to comment to TechCrunch, the details remain unconfirmed by the parties involved. Readers weighing the significance of the talks should treat the specifics — scope, cadence, participants — as reported claims rather than established fact until the companies describe them directly.