OpenAI, Anthropic and Google DeepMind Hold Private AI Safety Talks as Pressure Grows for Common Rules

OpenAI, Anthropic and Google DeepMind have been in private talks for weeks about AI safety, according to OpenAI policy chief Chris Lehane, as the industry faces mounting pressure to agree on guardrails for frontier models before they become more powerful and harder to control. The discussions matter because they signal rare coordination among leading rivals at a moment when lawmakers, regulators and some executives are debating whether the world’s most advanced AI systems need stricter oversight.

Lehane’s remarks on Tuesday, first reported by Bloomberg, add fresh evidence that the biggest AI labs are quietly exploring common standards, even as the companies continue competing aggressively on products, talent and market share. The timing is significant: the talks come just days after Anthropic chief executive Dario Amodei published a public call for the industry to slow down and build stronger safety mechanisms to reduce the chance of catastrophic harm.

Amodei’s essay, which landed on Saturday, framed frontier AI as a technology that could outpace society’s ability to supervise it. It urged companies and governments to work together on safeguards rather than waiting for a crisis. In the days since, prominent AI leaders have publicly echoed parts of that message, including OpenAI chief executive Sam Altman and Google DeepMind chief executive Demis Hassabis.

The result is an unusual convergence: public warnings about existential or catastrophic risk on one hand, and private cross-company coordination on the other. That combination has renewed questions about whether the AI industry can regulate itself, whether governments should step in, and whether some forms of cooperation could run into antitrust scrutiny if they are seen as limiting competition.

What did OpenAI confirm about the AI safety talks?

OpenAI confirmed that it has been speaking with Anthropic and Google DeepMind for several weeks about AI safety, according to Lehane, who serves as the company’s global policy chief. His comments amount to the clearest public acknowledgment yet that the three companies have been engaged in direct discussions on shared safety concerns.

Lehane was in Washington on Tuesday, where he was meeting with U.S. lawmakers focused on the possibility that highly advanced AI systems could produce severe or even catastrophic risks. That policy push suggests OpenAI is trying to shape the debate in Congress at the same time it is helping build consensus among major labs.

The confirmation also aligns with earlier hints from industry leaders. In a recent interview, Altman suggested that he had been in private conversations with other AI executives. Reporting from The Information this week further indicated that the companies had been discussing a possible standards body for the AI sector.

OpenAI’s policy chief said the company has been in talks with Anthropic and Google DeepMind for weeks about AI safety, underscoring that the major labs are exploring some form of coordination even while they compete in the market.

Why the timing matters now

The talks are happening during one of the most intense periods of debate over frontier AI. Models are improving quickly, companies are spending heavily to train them, and policymakers in the U.S. and abroad are trying to catch up with technical developments that move far faster than legislation.

At the same time, the conversation around AI risk has become more public and more urgent. Amodei’s essay was notable not just because it came from the chief executive of one of OpenAI’s closest competitors, but because it argued that the industry should intentionally slow the pace of frontier development if the risks warrant it. That is a striking proposal in a sector where speed is usually treated as a strategic advantage.

Support from Altman and Hassabis added weight to the message. When the leaders of OpenAI and Google DeepMind endorse a rival’s safety warning, it signals that concern about advanced AI is not limited to one company’s culture or product strategy. It is becoming part of the mainstream executive conversation.

How the industry is balancing competition and cooperation

The answer is that companies are trying to do both at once, but the balance is unstable. Rival labs still compete for users, enterprise contracts, cloud resources and top researchers. Yet when it comes to safety testing, model evaluation and possible industry standards, the incentives increasingly point toward cooperation.

That cooperation is not automatically straightforward. Some executives, including Altman, have acknowledged that direct coordination among major AI firms could invite antitrust concerns if it appears to suppress competition. In other words, the same conversations that could reduce risk might also draw legal scrutiny if regulators conclude the companies are cooperating too closely on market behavior.

Amodei attempted to address that tension in his essay by calling for a narrow government waiver that would permit safety coordination. Lehane, however, reportedly said the firms do not need such a waiver, suggesting OpenAI believes the talks can proceed under existing law if they remain focused on safety.

What are the companies actually discussing?

The full agenda has not been made public, but the reporting and statements so far point to a few likely themes: model testing, third-party evaluation, shared standards and possibly a formal body to coordinate frontier AI safety practices.

One concrete idea already on the table is the use of independent evaluators. Altman has said OpenAI would work with Anthropic on embedding third-party evaluators inside the company to assess safety. That would create an outside check on model behavior and development practices, rather than relying solely on internal testing teams.

The Information also reported that the companies have been discussing the creation of a standards body for the AI industry. Such a body could, in theory, define safety benchmarks, establish evaluation protocols and offer a venue for agreeing on when deployment of certain systems should be slowed or paused.

Hassabis has previously argued that the U.S. should create a new AI watchdog with the authority to review the most advanced models and coordinate industry-wide slowdowns if risks worsen. That idea resembles other safety regimes in high-stakes sectors such as aviation or pharmaceuticals, where independent oversight is considered essential.

Why are antitrust concerns entering the conversation?

Antitrust concerns are emerging because cooperation among direct competitors can sometimes be interpreted as an effort to restrain competition, even when the stated goal is public safety. That makes the AI safety discussions politically and legally delicate.

The issue is not hypothetical. If major labs agree on common deployment limits, testing rules or slowdowns, regulators could ask whether those measures protect the public or whether they also help incumbent firms reduce pressure from smaller rivals. The line between legitimate safety coordination and market coordination can be hard to draw.

Altman has reportedly recognized that risk, while Amodei’s call for a government waiver suggests he believes legal clarity may be needed before firms can coordinate safely and transparently. Lehane’s statement that a waiver is unnecessary implies OpenAI sees the current talks as narrow enough to avoid crossing that line.

That disagreement matters because it shapes what kind of cooperation is realistic. If companies and lawmakers cannot agree on legal protections, safety coordination may remain informal, limited or short-lived. If they can, the industry could move toward a more structured framework for frontier-model oversight.

How Washington fits into the picture

Washington is now central to the AI safety debate because the biggest questions may be too large for companies to settle alone. Lehane’s presence in the capital underscores that OpenAI is trying to influence legislative conversations as lawmakers examine catastrophic-risk scenarios.

One provision drawing support from OpenAI is included in the FRONTIER Act, which would require top frontier AI labs to allow “independent verification organizations” to inspect how advanced models are being built and tested. The idea is to create external oversight without waiting for a full federal regulatory agency to be built from scratch.

Such a measure could become part of a broader policy toolkit that includes reporting requirements, model evaluations, incident disclosure and emergency intervention powers. For companies, the appeal of a statutory framework is that it could reduce uncertainty and potentially allow more predictable safety collaboration.

But the political environment remains complicated. President Donald Trump has dismissed some AI safety warnings as exaggerated and argued that heavy regulation would hand an advantage to China. His AI adviser, David Sacks, has echoed that view, saying fears of existential AI risk are overstated.

That split in Washington could shape how far the industry’s safety push goes. If the White House and Congress remain skeptical of stricter oversight, companies may be left trying to build voluntary standards without a strong regulatory backstop. If lawmakers become more receptive, the current private talks could evolve into something more formal.

What is the FRONTIER Act?

The FRONTIER Act is a proposed U.S. measure aimed at increasing oversight of the most advanced AI systems. In this context, the legislation matters because it would create a mechanism for independent verification rather than relying only on internal company promises.

OpenAI’s reported support for the provision suggests the company is willing to accept some level of external scrutiny if it creates a clearer path for safe development. That position may help the company argue that it is serious about safety while also shaping the terms of any future regulation.

How did the public debate escalate so quickly?

The debate accelerated because the leading figures in AI have increasingly moved from abstract warnings to practical proposals. Amodei’s essay was a catalyst, but it landed on top of months of concern about model capability growth, agentic systems and the possibility that powerful models could be used in dangerous ways.

Once Altman, Hassabis and others publicly signaled agreement with parts of the call to action, the conversation shifted from one company’s warning to a broader industry theme. That matters because policy discussions often gain traction only when multiple major players acknowledge the same risk.

The disclosures also reveal a deeper reality: even the companies most responsible for pushing the frontier appear to believe that some problems may be too serious to solve through competition alone. If safety is a shared constraint, then common standards may become as strategically important as bigger models or better chips.

1. Anthropic’s chief executive publicly warned that frontier AI needs slower, more deliberate development.
2. OpenAI and Google DeepMind leaders signaled agreement with parts of that warning.
3. OpenAI later confirmed private talks with both rivals about safety.
4. Washington lawmakers are now being asked to consider oversight tools such as independent verification.

What happens next?

The most immediate next step is likely continued policy engagement in Washington and more behind-the-scenes coordination among the three companies. There is no sign yet that a formal industry body has been established, but the reporting suggests the groundwork is being laid.

If the companies move forward, they could define common evaluation standards, agree on the role of third-party testers and potentially recommend mechanisms for slowing deployment when certain danger thresholds are reached. If they stop short, the talks may still influence future legislation and shape public expectations around responsible AI development.

There is also a broader competitive question. Even if the companies can align on safety, they still have to decide how much openness to allow and how much strategic information to share. That tension may determine whether the talks become a durable feature of the industry or just a short-lived response to recent warnings.

For now, the clearest takeaway is that the AI race is entering a new phase. The leading labs are still competing, but they are also acknowledging that the next generation of models may require shared rules, outside verification and, possibly, government involvement to keep the technology from outrunning oversight.

Key figures and timeline

The recent sequence of events shows how quickly the safety debate moved from a single essay to direct inter-company talks and congressional engagement.

Bottom line

The confirmation that OpenAI, Anthropic and Google DeepMind have been in safety talks for weeks marks a notable shift in the AI industry’s public posture. The biggest labs are no longer only racing to release better models; they are also considering whether they need shared safeguards, outside evaluators and possibly a formal standards framework to manage the risks of frontier AI.

Whether those efforts become real policy, industry norms or another short-lived warning cycle will depend on the coming months in Washington and on whether the companies can cooperate without running afoul of competition law. For now, the message from the top of the AI sector is clear: the safety question is no longer theoretical.