OpenAI, Anthropic and Google DeepMind Have Spent Weeks Coordinating on AI Safety
OpenAI has spent weeks working with Anthropic and Google DeepMind on AI safety. Global policy chief Chris Lehane disclosed the coordination to reporters on September 15, 2026, according to TechCrunch.
TechCrunch said Bloomberg first reported Lehane's remarks, rather than reporting them itself. Reuters also reported the coordination on September 15, citing Bloomberg News, and named the participants as 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. Safety work among the leading labs is now tied directly to writing laws, not only to research posts or voluntary promises.
Coordination disclosed in Washington
Lehane described two policy details. First, he said OpenAI, Anthropic and Google DeepMind do not need a government waiver to coordinate on safety. That answered Anthropic CEO Dario Amodei, who had proposed a narrow waiver in an essay so labs could collaborate on safety without antitrust exposure, the legal risk of punishment for competitors working together. Lehane's position is that existing law already permits this type of coordination.
Second, Lehane said OpenAI supports a provision in the FRONTIER Act that would require the top frontier labs, the small group building the most powerful models, to let independent verification organizations check that models are developed safely. That would make outside safety checks a legal requirement, moving beyond voluntary evaluator programs.
Sam Altman had hinted at private coordination days earlier. In a Fortune interview published September 12, 2026, Altman said he had held private discussions with other AI leaders. Altman also said OpenAI would join Anthropic in embedding third-party evaluators to monitor safety. That gives outside reviewers ongoing access inside development, rather than limited API access to test a finished model once before release.
A standards body and third-party verification
The Information reported on September 13, 2026 that OpenAI, Anthropic and Google DeepMind have been working together to create a standards body for the AI industry, according to The Information. The report described private discussions about an industry-led system to oversee development practices.
A useful comparison is building codes and building inspectors. A standards body sets testable expectations for pre-deployment evaluation, incident reporting, and secure handling of models. Embedded evaluators check whether those expectations hold in working systems. The FRONTIER Act provision points to the second job. The reported industry body points to the first.
Precedent and pressure
Joint safety work among these labs is not new. On June 12, 2023, OpenAI, Google DeepMind and Anthropic agreed to provide early or priority access to models to support UK AI safety research. On July 26, 2023, OpenAI, Microsoft, Google and Anthropic formed a new body to support safe and responsible frontier AI development. OpenAI and Anthropic later tested each other's models for misalignment in a first-of-its-kind joint safety evaluation, 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 to defend against rogue AI. On July 30, 2026, Anthropic said its own AI models had breached three companies during security tests. On July 16, 2025, OpenAI and Anthropic researchers criticized the safety culture at Elon Musk's xAI as reckless.
Anthropic published a threat-intelligence report in September 2026. It covers threat actors disrupted between December 2025 and August 2026 across seven areas of harm, from cyber operations to biological misuse. The scope includes state-linked intrusion work and dual-use uplift, where general AI abilities could help a bad actor cause more harm, the areas where evaluations, access controls, and disclosure rules are tested most directly.
The broader context here will be familiar to anyone who followed cloud security or mobile platform review. Competing vendors can share threat signals and test methods without sharing model weights, training data, or product plans. Antitrust caution is real, but there is precedent for narrow technical cooperation on abuse detection and incident response.
In my view, the question that will last is verification. Voluntary model access helped start government safety institutes. Persistent evaluator access and independent verification would change daily work. Labs would need logging, reproducibility, and version pinning so an outsider can confirm the tested model is the same as the deployed system. That is solvable engineering. Worth flagging here is that it would shift safety from descriptions of process to evidence about specific builds.
Stepping back, the longer arc here stays constructive. If the weeks-long talks produce a common evaluation interface and a shared incident taxonomy, a shared list of failure categories, both enterprise buyers and regulators gain. Developers would face fewer one-off safety reviews per deployment. Researchers could compare results across labs. Better measurement rarely slows useful technology for long. It tends to make wider deployment easier to defend.
