The Biggest AI Labs Are Teaming Up on Safety
OpenAI has been working with Anthropic and Google DeepMind on AI safety for weeks. Policy chief Chris Lehane told reporters about the talks on September 15, 2026, according to TechCrunch.
TechCrunch said Bloomberg first reported Lehane's remarks. Reuters also reported the talks on September 15, citing Bloomberg News, and named OpenAI, Anthropic and Alphabet's Google DeepMind.
Lehane was speaking in Washington, where he said he was working with U.S. lawmakers on catastrophic risks from AI.
The broader context here starts with Washington. The safety talks are now connected to writing laws, not just research posts or voluntary promises.
Coordination disclosed in Washington
Lehane made two policy points. First, he said the three labs do not need special government permission to work together on safety. That answered Anthropic CEO Dario Amodei, who had suggested a narrow legal waiver in an essay so labs could cooperate without breaking competition laws. Lehane said current law already allows it.
Second, Lehane said OpenAI supports part of the FRONTIER Act. That part would require the leading labs building the most powerful models to let independent checkers verify that models are built safely. That would make outside checks required by law, not just voluntary.
Sam Altman had hinted at private talks days earlier. In a Fortune interview published September 12, 2026, Altman said he had spoken privately with other AI leaders. He also said OpenAI would join Anthropic in giving outside reviewers ongoing access for safety, instead of letting them test a finished model once through a limited connection.
A standards body and third-party verification
The Information reported on September 13, 2026 that OpenAI, Anthropic and Google DeepMind have been discussing a shared standards group for the AI industry, according to The Information. The report described private talks about an industry-run way to oversee how AI is built.
Think of building inspectors. One job is to write the safety checklist. The other is to visit the site and check the work. The standards group would set the checklist for testing before launch, reporting problems, and handling models safely. Outside reviewers would do the checking. The FRONTIER Act part is about the checking. The reported group is about the checklist.
Precedent and pressure
This teamwork is not new. On June 12, 2023, OpenAI, Google DeepMind and Anthropic promised early or priority access to models for UK safety research. On July 26, 2023, OpenAI, Microsoft, Google and Anthropic formed a group to support safe frontier AI work. OpenAI and Anthropic later tested each other's models for misalignment, dangerous behavior where a model acts against its intended goals, with results published August 27, 2025.
Three recent events came before this renewed coordination. On August 27, 2026, OpenAI, Anthropic, Google and more than 100 other companies called for action against rogue AI. On July 30, 2026, Anthropic said its own models had broken into three companies during security tests. On July 16, 2025, OpenAI and Anthropic researchers called the safety culture at Elon Musk's xAI reckless.
Anthropic published a threat-intelligence report in September 2026. It covers bad actors stopped between December 2025 and August 2026 across seven harm areas, from hacking to biological misuse. It includes government-linked hacking and dual-use uplift, where general AI skills could help someone cause more harm, where testing, access limits, and reporting rules matter most.
The broader context here is one I have seen before with cloud security and phones. Rival companies can share warnings about attacks and how they test for them, without sharing their core technology or business plans. Care about competition law is real, but there is past practice for limited teamwork on abuse and incident response.
In my view, the lasting question is proof. Voluntary access helped start government safety groups. Ongoing access for outside checkers would change daily work. Labs would need clear logs, repeatable tests, and version control so a checker can confirm the tested model is the one actually released. That can be built. Worth flagging here is that it moves safety from promises about process to evidence about finished models.
Stepping back, the longer arc here stays hopeful. If the talks lead to shared tests and a shared list of failure types, both business users and regulators benefit. Developers would face fewer custom safety reviews. Researchers could compare results across labs. Better measurement tends to make wider use easier to support.
