OpenAI Confirms Weeks of Secret AI Safety Talks With Anthropic and Google, as White House Dismisses Risk Concerns

If you've ever wondered whether the companies building the world's most powerful AI systems actually talk to each other about the risks, we now have an answer. They do. Quietly. For weeks.

Chris Lehane, OpenAI's global policy chief, told reporters on Tuesday that his company has been in ongoing discussions with rivals Anthropic and Google DeepMind about AI safety, confirming reporting first published by Bloomberg. Lehane made the admission while in Washington, where he's been meeting with lawmakers about what the industry now euphemistically calls "catastrophic risk," the possibility that an AI system causes harm on a scale we can't easily undo.

Think of it like three competing airlines quietly agreeing to share turbulence data, even as they fight over ticket prices. The cooperation makes sense from a safety standpoint. It also raises uncomfortable questions about what happens when your competitors become your safety board.

The talks follow an essay published Saturday by Anthropic CEO Dario Amodei, who called on the industry to slow the pace of frontier AI development and work collectively to avoid worst-case outcomes. That's a notable ask from a company whose business model depends on staying at the front of that same race. The response was swift and, for an industry known for public sniping, unusually unified. OpenAI's Sam Altman, Google's Demis Hassabis, and even SpaceX and xAI's Elon Musk all voiced support on social media. Altman went further, saying OpenAI would join Anthropic in bringing in third-party evaluators to monitor its own safety practices from the inside.

That's a meaningful shift. Third-party evaluators are essentially independent auditors, people outside the company whose job is to check whether the safety claims a lab makes about its own models hold up under scrutiny. Letting outsiders into your building to check your homework isn't something AI labs have historically been eager to do.

The Antitrust Tightrope

Here's where it gets complicated. When competitors coordinate on anything, even something as seemingly uncontroversial as safety, antitrust law starts paying attention. Altman himself has flagged that risk, noting that safety talks between rival firms could be read as suppressing competition if they're not handled carefully. Amodei's essay anticipated this problem and proposed a narrow government waiver that would let companies coordinate on safety without running afoul of the law.

Lehane, notably, said OpenAI doesn't think it needs one.

That's a confident legal position to take on an unsettled question, and it's worth watching whether regulators agree. The stakes go beyond process. The Information reported this week that the three companies have been working to build a standards body for the AI industry, essentially a shared rulebook for what safe development looks like. Altman reportedly told OpenAI staff this effort would have to move forward without support from the U.S. government, a striking admission from a company that has spent years courting Washington.

It's also not a new idea. Hassabis called for something similar back in July, pushing for a government-backed watchdog with the authority to screen the most advanced AI models and, if needed, coordinate an industrywide slowdown. The fact that industry insiders are now building a private version of that watchdog, rather than waiting for one from government, tells you something about where they think the political winds are blowing.

That reading gets confirmation from the very top of the U.S. government. President Trump has dismissed AI safety concerns outright, reportedly calling them a "hoax" and framing any push for tighter regulation as a gift to China in the global AI race. His AI advisor David Sacks, a longtime tech investor with his own financial stake in the industry, echoed that sentiment, calling fears of existential AI risk overblown.

That leaves the safety-minded wing of the AI industry in an odd position: pushing forward with self-regulation not because Washington asked them to, but because Washington seems actively uninterested in asking. Companies that in another political climate might have preferred clear government rules over ambiguous self-policing are now building their own guardrails and hoping antitrust regulators don't call foul.

Not every lawmaker is ignoring the issue. At the same Tuesday briefing, Lehane said OpenAI supports a provision in the bipartisan FRONTIER Act that would require top AI labs to allow "independent verification organizations" to review whether their models are being developed safely. It's a smaller, more targeted step than a full industry watchdog, but it suggests at least some appetite in Congress for outside oversight, even if the White House isn't pushing for it.

Why It Matters

The public rarely gets to see how the companies building AI systems that could reshape work, information, and daily life actually think about risk behind closed doors. This moment offers a rare glimpse, and it's a mixed picture. On one hand, fierce competitors are choosing to cooperate on something that matters more than market share. On the other, they're doing it quietly, without clear government backing, and in a legal gray zone that nobody has fully mapped out.

For everyday people, the practical question isn't whether these companies are talking. It's whether that talking translates into real safeguards, verified by someone other than the companies themselves, before the next generation of AI systems reaches the public. Voluntary cooperation among rivals can be a meaningful first step. It's not, on its own, a substitute for enforceable rules, especially when the political leadership overseeing the industry seems more worried about losing a race than about what happens if everyone crosses the finish line too fast.