OpenAI, Anthropic and Google Talk Safety Standards
OpenAI's policy chief confirms weeks of safety coordination with Anthropic and Google DeepMind, as the White House plays the risks down.
In short
OpenAI, Anthropic and Google DeepMind have spent several weeks coordinating on shared safety testing for frontier models, OpenAI policy chief Chris Lehane confirmed on September 15, 2026.
At a glance
- OpenAI policy chief Chris Lehane confirmed weeks of safety talks with Anthropic and Google DeepMind on September 15, 2026.
- The plan centers on an industry-funded body that would test frontier models before public release.
- Demis Hassabis floated such a body in July; Sam Altman wants outside evaluators embedded in OpenAI teams.
- Antitrust remains unresolved: Amodei raised a government waiver, Lehane says none is needed.
- Cohere CEO Aidan Gomez calls the effort a cartel dressed up as safety.
OpenAI has put a name to something the industry had kept quiet. Policy chief Chris Lehane said on September 15, 2026 that the company has spent several weeks coordinating with Anthropic and Google DeepMind on how frontier models should be tested for danger. The trigger was a weekend essay by Anthropic CEO Dario Amodei calling on rivals to slow the pace of frontier development together. Sam Altman, Demis Hassabis and Elon Musk publicly backed the appeal.
What is on the table
The centerpiece is an industry-funded standards body that would evaluate powerful systems before they reach the public. Reports describe the model as analogous to the private self-regulator that oversees U.S. brokerage firms. Demis Hassabis proposed such a body in July, arguing it would need serious money to attract researchers and buy compute for large-scale testing. Altman has signaled that OpenAI would bring independent evaluators inside its own development teams.
Antitrust is the sticking point
Competitors agreeing on how fast to move is exactly the shape of conduct antitrust law scrutinizes. Amodei suggested the government could grant an explicit waiver. Lehane disagreed, saying companies do not need permission to cooperate on safety. No agency has weighed in either way.
Washington is pulling in two directions
President Donald Trump has dismissed the warnings as a hoax and points to the race with China; advisor David Sacks calls talk of existential risk exaggerated. Congress is moving the other way. The FRONTIER Act would require independent verification organizations inside frontier labs, and a separate bill would hand the Commerce Secretary authority over safety audits. Senator Bernie Sanders rejects voluntary pledges outright and wants binding international rules.
The cartel charge
Cohere CEO Aidan Gomez describes the initiative as a cartel operating under the banner of safety. Nvidia CEO Jensen Huang has called the safety alarm irresponsible. Criticism also comes from the opposite direction: former researcher Jacob Coxon accused the labs of gambling with everyone's lives.
What is not settled
The substance of the talks has not been made public. None of the three companies has named a timeline, a legal structure or a budget for the proposed testing body. Whether any of this becomes an institution with real authority cannot be established from what is on the record — the reporting describes conversations, not a signed agreement.
FAQ
What exactly are OpenAI, Anthropic and Google DeepMind discussing?
Shared safety standards and a possible testing body that would evaluate frontier models before release. The specifics of the talks have not been published.
Can rival AI labs legally coordinate like this?
That is unresolved. Amodei raised the idea of a government waiver, Lehane says none is required, and no regulator has taken a position so far.
Has anything actually been agreed?
No. What is confirmed is several weeks of talks plus statements of intent; the timeline, legal form and funding of the testing body remain open.
