OpenAI, Anthropic, and Google Are Quietly Coordinating on AI Safety

OpenAI, Anthropic, and Google DeepMind have been quietly talking to each other about AI safety for weeks, and the trigger was a single essay from Anthropic’s CEO. On September 15, 2026, OpenAI’s global head of policy Chris Lehane confirmed to reporters that the three rival labs have been coordinating on a shared safety framework since Anthropic CEO Dario Amodei published “We Must Pace the Frontier” on September 12. I’ve read a lot of AI-safety open letters over the past few years that amounted to nothing more than a press cycle. This one is different, because it produced an actual admission from a competitor’s policy chief that back-channel talks were already underway before the essay even went public.

Here’s what happened, why it matters, and what it actually commits any of these companies to do — which, as of this week, is less than the headlines suggest.

What Amodei’s essay actually argues

Amodei’s roughly 3,800-word piece, posted to his personal site rather than Anthropic’s blog, makes a narrower claim than “stop building AI.” He writes that the industry needs to deliberately slow the rate at which it increases model capability — not halt training runs, not freeze research, but stretch out the interval between capability jumps so that alignment work, third-party verification, and basic operational safety can catch up. His own phrase for it is a “speed limit,” not a stop sign.

The risks he lists for moving too fast are specific: loss of control over autonomous systems, AI-assisted cyberattacks, bioterrorism uplift, economic disruption from labor displacement happening faster than institutions can adapt, and — the most eyebrow-raising line in the essay — a scenario where autonomous bots effectively take over large parts of “the entire internet,” which he estimates could plausibly start within six to twelve months if agentic systems keep scaling unchecked.

The China problem nobody has an answer for

Amodei didn’t dodge the obvious objection. In a follow-up on September 13, he called out what he termed the “toughest dilemma” in his own proposal: if the U.S. labs voluntarily throttle their own capability gains and Chinese labs don’t reciprocate, a unilateral slowdown just hands the frontier to whoever ignores the speed limit. His stated fix isn’t a pledge — it’s an aspiration to eventually negotiate a mutual pace agreement, similar in spirit to arms-control verification regimes, though he’s explicit that nothing like that exists yet.

That gap between “we should coordinate internationally” and “we have coordinated internationally” is exactly where this story sits right now. What’s confirmed is domestic industry talk, not a global treaty.

Who actually signed on, and to what

The response from other lab leaders was fast and public, but the commitments vary a lot in weight:

• Sam Altman (OpenAI) backed the essay and said OpenAI would embed independent third-party evaluators directly inside the company to monitor frontier model development — a concrete, if vague-on-details, operational commitment.
• Demis Hassabis (Google DeepMind) had already floated a separate idea before Amodei’s essay: a new standards body to screen advanced models before wide release. His support folds that proposal into the same conversation.
• Elon Musk voiced support for the letter through SpaceXAI, without committing his own frontier efforts to anything specific.

Lehane’s addition on September 15 was the part that turned this from “CEOs post supportive quotes” into an actual story: he said the three companies don’t think they need government sign-off to coordinate this way, despite the antitrust questions that competitors jointly setting technical standards inevitably raises. He also confirmed OpenAI supports provisions in the pending FRONTIER Act that would require independent verification organizations to audit frontier labs — which is a notably different posture than treating safety coordination as purely voluntary and self-policed.

Washington isn’t on board

The Trump administration’s reaction undercuts the “industry unites for safety” framing a little. The White House dismissed the catastrophic-risk concerns as overblown, and AI advisor David Sacks argued that any regulatory slowdown in the U.S. would simply hand the advantage to China — the exact tension Amodei himself flagged as unresolved. That leaves the three labs coordinating largely on their own initiative, with an administration that isn’t asking for it and a legislative process (the FRONTIER Act) that’s still pending, not passed.

Why this is worth taking seriously anyway

I’d normally file “rival AI CEOs voice support for a safety essay” under performative. What moves this past that is the sequencing: the informal talks predate the essay’s publication, meaning this wasn’t three competitors reacting to pressure — it was a coordinated move where Amodei’s public essay gave cover to conversations that were already happening privately. That’s a meaningfully different governance pattern than the open-letter era of 2023, and it lines up with the broader pattern this year of frontier labs treating safety incidents — like GPT-6 Astra crossing the “Critical” cybersecurity threshold under OpenAI’s own Preparedness Framework — as things serious enough to slow down for, not just disclose after the fact. It also follows scientists at September’s UN Global Dialogue on AI Governance telling governments outright that no technical guarantee exists that advanced AI avoids catastrophic harm — the same warning Amodei is now making from inside industry rather than from a UN panel.

It also arrives in the same stretch as governments moving on their own tracks — the EU’s DSA designations, state-level AI acts, and now a federal push described in what the Great American AI Act would actually do — which makes an industry-led standards body look less like altruism and more like labs trying to shape the rules before regulators write them without industry input.

What to watch next

Three things will tell you whether this becomes real: whether OpenAI’s third-party evaluators get named and given actual access rather than a symbolic role, whether Hassabis’s proposed standards body gets a charter and funding, and whether the FRONTIER Act’s audit requirements survive the legislative process largely intact. Until at least one of those lands, this is a coordination announcement, not a policy.

FAQ

What did Dario Amodei actually propose?

In his September 12 essay “We Must Pace the Frontier,” Amodei argued the AI industry should deliberately slow the rate of capability improvement — not stop building models — to give alignment research, third-party verification, and safety operations time to keep pace with what frontier models can do.

Are OpenAI, Anthropic, and Google actually working together now?

OpenAI’s Chris Lehane confirmed on September 15 that the three companies have been in informal talks on a shared safety framework for several weeks, predating Amodei’s public essay. No joint body or binding agreement has been finalized yet.

Does this require government approval?

Lehane said the companies believe they can coordinate on safety standards without needing government sign-off, despite the antitrust questions raised by competitors jointly setting technical standards.

What is the FRONTIER Act?

It’s pending U.S. legislation that would require independent verification organizations to audit frontier AI labs. OpenAI has said it supports the act’s audit provisions, giving the informal industry talks a potential legislative backstop.

Why does China complicate the plan?

Amodei himself acknowledged that a voluntary U.S. slowdown only works if rival nations, particularly China, don’t use the pause to close the capability gap — a dilemma he says has no current solution beyond an aspirational future negotiation.