Google, OpenAI and Anthropic are pushing forward with a plan to create an independent, industry-led body to set safety standards for frontier AI, The Information reported on September 24, citing unnamed sources. The target launch window: the end of this year or early next year.
If it ships, the three companies building the most capable AI systems on the planet would share one set of evaluation protocols for their own products — a structure with no real analogue in any other safety-critical industry.
What is actually on the table
The three labs have been in working-group discussions since July 2026. The goal, per reporting: shared protocols for evaluating frontier models, independent safety assessments, and pre-release checks before systems reach the public. A tentative name floating in earlier reporting was the “Frontier AI Standard Agency.” Governance details — who chairs it, who funds it, what happens when a member fails its own standard — remain unresolved months into the talks.
The plan has a recent lineage. OpenAI and Anthropic had been negotiating a legally binding agreement to test each other’s commercially available models for vulnerabilities, with neither company retaining data from the tests. Demis Hassabis proposed a public-private partnership with federal oversight, modelled on FINRA, on September 15. Anthropic CEO Dario Amodei made the intellectual case in a September 12 essay calling for external evaluators to sit at the centre of AI safety. And OpenAI’s global policy chief Chris Lehane confirmed the coordination is real, while rejecting the idea that it needs antitrust exemptions.
The credibility problem writes itself
Here is the tension, and it is not subtle. All three companies scored C+ or lower on the 2026 Future of Life Institute AI Safety Index, which grades frontier developers on transparency, safety implementation and responsible development practices. The companies proposing to certify frontier AI safety are, by a third-party rubric, the sector’s underperformers.
There is also history. The same three companies co-founded the Frontier Model Forum in 2023 — a safety initiative that, two years on, has no enforcement record anyone can point to. Each lab also maintains its own safety framework: Anthropic’s Responsible Scaling Policy ties model deployment to internal benchmarks, and Google DeepMind and OpenAI run internal evaluation pipelines. The proposed body would nominally sit above all of that. Whether it holds authority over members or merely documents them is the entire question, and it is unanswered.
The vacuum it is filling is real
The counterweight to the cynicism: governments are not filling this space. In the US, the White House has moved from a revoked safety order to a downsized voluntary-review framework, and told the UN that lab calls for common standards overreach — while Altman and Amodei were standing in the same building asking for them. The EU’s AI Act has enforcement powers, but its high-risk provisions have been delayed to 2027. Twenty-two countries signed a call for frontier AI to stay under human control at UNGA week with no mechanism attached.
Into that gap step three companies with an obvious conflict of interest and an obvious capability advantage. Nobody else can evaluate a frontier model right now — the talent and the compute sit inside those buildings. An industry-run standard with real teeth beats a government standard that exists on paper. An industry-run standard that merely brands itself as safety is worse than nothing, because it manufactures the appearance of oversight.
Which of those two things this body becomes depends on details its founders have not agreed to yet: independent funding, published results, and the right to fail a member’s model. The reporting suggests the shape exists. The teeth do not.
