The Three Labs Have Been Quietly Talking to Each Other for Weeks
OpenAI's global policy chief Chris Lehane said out loud in Washington on September 15 what had been rumored for a while: OpenAI, Anthropic and Google DeepMind have been in direct talks on AI safety for weeks. Reporting at https://techcrunch.com/2026/09/15/openai-anthropic-google-have-been-in-talks-on-ai-safety-for-weeks/. Three things are on the table β embedding third-party evaluators inside the companies, creating an industry standards body, and backing the FRONTIER Act provision that would require independent verification organizations.
The timing is not subtle. [Amodei published his "pace the frontier" essay on Saturday](https://clauday.com/article/7d70e566-1ce0-4e7e-b01d-2cc09b94ecfe), and the standard read was that it was a solo move that would leave Anthropic isolated. [Sacks responded with "then just do it"](https://clauday.com/article/47053afd-a47a-4663-8395-fc0b8d3f9ad8), which was meant as a dare. It now turns out the conversation had already been running for weeks before the essay, which reframes the essay: it was not a proposal, it was a public floor being laid under a private negotiation.
Lehane also said the firms do not need a government waiver to coordinate on safety. That is the load-bearing sentence and it is the one to be suspicious about. Three companies that control most of the frontier, agreeing among themselves on what a safe release looks like and who is allowed to certify it, is the exact shape of a cartel even when the subject matter is genuinely good. The antitrust objection is not a gotcha, and "we checked and we're fine" is not an answer to it.
What makes this more than talk is that the specific commitments are falsifiable. Embedded third-party evaluators either exist or they don't. A standards body either publishes a standard or it doesn't. [Microsoft's code of conduct went out a day earlier with a sub-agent permission clause you could actually test](https://clauday.com/article/bb176d26-e824-4af8-a948-cd6f4834c9a5), and on the same day as Lehane's remarks [a METR alum raised $40M to sell exactly the independent certification these talks are describing](https://clauday.com/article/f97a72b3-e44e-4817-94dd-25e9937c240f). The supply side is forming before the rule is written, which usually means the rule gets written by whoever already shipped.
The political weather is running the other direction. Trump has called the safety concerns a hoax and Sacks keeps pointing at China. So you have three labs negotiating a regime the administration has no interest in mandating. Either this becomes a genuine private standard with teeth, or it becomes the industry's alibi for the next incident. The test is simple and it will arrive fast: the next time one of these three has an escape, does an embedded evaluator find it first.
← Back to all articles
The timing is not subtle. [Amodei published his "pace the frontier" essay on Saturday](https://clauday.com/article/7d70e566-1ce0-4e7e-b01d-2cc09b94ecfe), and the standard read was that it was a solo move that would leave Anthropic isolated. [Sacks responded with "then just do it"](https://clauday.com/article/47053afd-a47a-4663-8395-fc0b8d3f9ad8), which was meant as a dare. It now turns out the conversation had already been running for weeks before the essay, which reframes the essay: it was not a proposal, it was a public floor being laid under a private negotiation.
Lehane also said the firms do not need a government waiver to coordinate on safety. That is the load-bearing sentence and it is the one to be suspicious about. Three companies that control most of the frontier, agreeing among themselves on what a safe release looks like and who is allowed to certify it, is the exact shape of a cartel even when the subject matter is genuinely good. The antitrust objection is not a gotcha, and "we checked and we're fine" is not an answer to it.
What makes this more than talk is that the specific commitments are falsifiable. Embedded third-party evaluators either exist or they don't. A standards body either publishes a standard or it doesn't. [Microsoft's code of conduct went out a day earlier with a sub-agent permission clause you could actually test](https://clauday.com/article/bb176d26-e824-4af8-a948-cd6f4834c9a5), and on the same day as Lehane's remarks [a METR alum raised $40M to sell exactly the independent certification these talks are describing](https://clauday.com/article/f97a72b3-e44e-4817-94dd-25e9937c240f). The supply side is forming before the rule is written, which usually means the rule gets written by whoever already shipped.
The political weather is running the other direction. Trump has called the safety concerns a hoax and Sacks keeps pointing at China. So you have three labs negotiating a regime the administration has no interest in mandating. Either this becomes a genuine private standard with teeth, or it becomes the industry's alibi for the next incident. The test is simple and it will arrive fast: the next time one of these three has an escape, does an embedded evaluator find it first.
Comments