OpenAI says it is working with Anthropic and Google on AI safety

Three companies that compete on every model release have spent several weeks discussing something none of them can build alone: one organization with the standing to test their systems before those systems ship.

What was confirmed

Chris Lehane, OpenAI’s global policy chief, described the talks at a Washington briefing on September 15, 2026, and said the engagement with Anthropic and Google had been under way for several weeks. The model under discussion is a self-regulatory organization along the lines of the Financial Industry Regulatory Authority, the body that oversees US brokers, adapted to test the most powerful AI systems before they reach the public.

The proposal originated with Demis Hassabis, who called for such an organization in July and argued that funding would need to be substantial and likely mostly industry-paid, in order to attract technical talent and provide compute for large-scale testing. Lehane said OpenAI does not see the need for an antitrust waiver to coordinate on safety, citing the airline industry as precedent. Amodei’s essay takes the opposite view, arguing that regulators do need to issue a narrow waiver for certain safety conversations.

One limit belongs on the record. Spokespeople for Anthropic and Google DeepMind did not immediately comment, so this is OpenAI’s account of where all three companies stand.

The warning behind it

The talks became public days after Amodei published his essay arguing the industry must pace the frontier. In it he wrote that his worry is that within six to twelve months an agent swarm could be capable of taking over the entire internet with a persistent botnet, potentially causing hundreds of billions of dollars in damage. The essay links an investigation report on an agent-driven attack as the basis for that forecast.

What would make it real

Self-regulation written by the regulated has a structural flaw, and Congress is moving separate bills in parallel. The detail that decides whether this body is more than a press position is access: an evaluator with employee-level reach into systems before release can check claims that no outside party can check today, while one that reviews what the labs choose to hand over cannot.