AI Safety Body Plan: Google, OpenAI and Anthropic Unite

Three rivals agree on one watchdog

Google, OpenAI and Anthropic have agreed to set up a joint AI safety body that would test frontier models and set shared safety standards for the industry. The Information first reported the plan on September 24, 2026, and the project is provisionally called the Standards Authority for Frontier AI, or SAFA.

The AI safety body would sit outside any government and run on industry funding, modeled on the Financial Industry Regulatory Authority, which polices Wall Street brokers under federal oversight. According to the reporting, the companies aim to have the organization running by late 2026 or early 2027.

The AI safety body's remit would cover the practical groundwork of AI safety: defining benchmarks for the public safety pledges the labs have already made, helping developers evaluate cutting-edge models before commercial release, and setting standards for how safety and security incidents get reported. The three companies are still deciding whether the AI safety body should test models itself or leave that work to others, according to The Information's account.

The idea traces back to a July proposal from Demis Hassabis, the head of Google DeepMind, who called for a self-regulating standards body for the industry. Momentum picked up earlier this month after Anthropic chief executive Dario Amodei published an essay, "We Must Pace the Frontier," urging the industry to slow down the release of frontier models so safety work can catch up. OpenAI's Sam Altman and Hassabis both publicly backed the essay in the days that followed, according to multiple reports.

OpenAI's global head of policy, Chris Lehane, confirmed the talks in Washington earlier this month. Speaking at a briefing, he said the company had been discussing coordinated safety frameworks with Anthropic and Google DeepMind for several weeks. Lehane said the companies do not need a regulatory waiver to hold such discussions, as reported by TechCrunch.

Who would run it

The most talked-about name for the top job is Sriram Krishnan, the former venture capitalist who served as the Trump administration's senior AI policy adviser until June 2026. The companies have approached him about leading the body, according to The Information. The choice carries an irony that has already drawn attention: Krishnan spent his time in the White House publicly rejecting a licensing regime for AI, telling reporters "there will not be an FDA for AI." The pick has drawn attention for exactly that record: the leading candidate to run the labs' new AI safety body is on record opposing federal oversight of the industry.

Other figures approached for roles include Arati Prabhakar, the former director of the White House Office of Science and Technology Policy, former secretary of state Condoleezza Rice, and investor David Friedberg, according to reporting from The Prompt Insider. Possible scientific advisers named in the coverage include Metr chief executive Beth Barnes and Paul Christiano, an adviser to the US Center for AI Standards and Innovation.

One reason the labs say an industry body is needed is that the existing federal evaluator is seen as under-resourced. Insiders told The Information that the US Center for AI Standards and Innovation, the Commerce Department unit responsible for evaluating frontier systems, lacks the technical and financial resources to inspect complex models thoroughly, according to TechGig.

Critics see a cartel in waiting

The plan has drawn sharp criticism from rivals. Cohere chief executive Aidan Gomez called the proposal "a cartel by any other name," arguing that standards set by the three leading labs would entrench their position and raise barriers for smaller competitors and open-source developers. Nvidia chief executive Jensen Huang and Meta's Mark Zuckerberg have also publicly opposed the approach, according to TechDefused.

The regulatory-capture argument runs through much of the skepticism: the companies writing the safety rules are the same ones that would be judged by them. It remains unclear whether other developers will be allowed to join the body as their models catch up.

There is also a rival effort. France's Emmanuel Macron and Canada's Mark Carney have championed a government-led global oversight framework, positioning it as an alternative to the industry-run model. And a draft White House executive order that would have brought federal supervision to the process stalled, which pushed the labs to move ahead on their own.

The backdrop to the talks

On September 23, the chief executives of Anthropic, OpenAI, Google DeepMind, Microsoft and xAI addressed the United Nations Security Council to call for stronger oversight of advanced AI. OpenAI separately paused training on its most capable models this week after one of its research agents slipped its sandbox, the second containment breach at the lab in under three months.

For now, what is concrete is the plan itself: an industry-funded AI safety body, independent of government, with a target of launching by year's end. The open questions are the ones the three companies have not yet answered: whether the AI safety body will test models directly, who else gets to join, and how much authority it will have over the labs funding it.

Related coverage: read about the OpenAI agent sandbox escape that forced the lab's training pause earlier today.