Anthropic, OpenAI and Google reportedly discussed forming AI safety body

The report adds detail to the safety-related statements from AI industry leaders that appeared over the weekend, indicating that coordination has progressed beyond a single essay or public declaration. In financial terms, a credible move toward shared testing and auditing frameworks could be interpreted in two different ways. Standardised pre-release safety checks could lower regulatory risk for the biggest labs by staying ahead of government action, a pattern that has in the past boosted valuations throughout the sector. However, several observers have pointed out that the costs of complying with a lab-designed standards body could strengthen the position of the largest incumbent firms, given that the OECD has already identified high fixed costs and concentrated infrastructure as obstacles to entering the AI market, a development to monitor for its impact on competition among smaller players and infrastructure providers.

The most prominent safety statements from the industry over the weekend turned out to be based on months of quiet, coordinated discussions among the three major labs.

Working group meetings involving Anthropic, OpenAI and Google have taken place since July to discuss the creation of an industry-driven AI safety standards body, according to The Information, which cited sources familiar with the talks. The discussions reportedly began before Anthropic chief executive Dario Amodei's public request on Saturday for AI companies to reduce the speed of capability advancements and improve safety coordination.

The reporting indicates that the public statements made over the weekend, including Amodei's essay and OpenAI chief executive Sam Altman's public endorsement of it, were based on an already active private coordination effort rather than starting one. Google DeepMind chief executive Demis Hassabis had previously set some of the foundation, proposing on July 14 a US-based Frontier AI Standards Body modelled on the Financial Industry Regulatory Authority, the self-regulatory organisation that oversees US broker dealers.

During an internal all-hands meeting, Altman reportedly stated that he supports the creation of a testing and auditing body for the AI industry, while noting that without US government support, the main AI labs would need to develop such a body themselves. The three companies seem to hold somewhat differing perspectives on the appropriate mix of industry self-regulation and government participation. Anthropic has tended more toward collaboration with government agencies, with the company's red team lead having publicly argued in July for safety standards co-developed with regulators. OpenAI, meanwhile, has stressed the importance of pushing forward voluntary standards independent of federal action, while also supporting specific state-level legislation regarding AI oversight.

The main proposal being discussed involves shared protocols for testing frontier models before they are released to the public, including independent evaluations, pre-launch safety reviews, and standardised risk assessments, based on reporting about the talks. Each company already has its own internal system for such assessments, with Anthropic's Responsible Scaling Policy and OpenAI's Preparedness Framework both outlining internal procedures for evaluating catastrophic risk before deployment, while Google DeepMind has been relatively less open about the details of its own internal safety procedures.

This is not the first effort at cross-company AI safety collaboration. The Frontier Model Forum, an industry-backed nonprofit started in 2023 by six major AI companies, operates an associated AI Safety Fund with more than $10 million allocated to independent safety research. More recently, the Linux Foundation's Agentic AI Foundation, formed in December 2025, brought together open standards projects such as Anthropic's Model Context Protocol, Block's Goose, and OpenAI's AGENTS.md. Whether the ongoing talks result in a more formal standards body, and how much authority or government support it eventually holds, remains unclear, with reporting indicating that the working group discussions are still in progress.