OpenAI, Anthropic and Google Explore Joint AI Safety Standards Body

OpenAI, Anthropic and Google are discussing the possibility of creating an industry-led body focused on standards and safety testing for increasingly powerful artificial intelligence systems.
The discussions reflect a growing question inside the AI industry: as frontier models become more capable, should companies develop common methods for testing and evaluating them rather than relying entirely on separate internal safety processes?
The talks are reportedly still at the discussion stage, meaning there is no final organisation, charter or regulatory authority that has been formally established from the discussions.
The FINRA Connection
The idea has been compared with the Financial Industry Regulatory Authority, commonly known as FINRA, a US organisation that oversees aspects of the securities industry.
Google DeepMind CEO Demis Hassabis previously proposed an AI standards body modelled partly on that approach.
The idea is to create an organisation that could bring together technical experts and industry participants to develop common methods for assessing advanced AI systems.
The proposal comes as AI developers face increasing pressure to demonstrate that increasingly capable models are tested for dangerous or unexpected behaviour.
Why Common Testing Matters
AI companies currently conduct their own evaluations and publish different kinds of safety information.
Although these internal evaluations can be useful, comparing results across companies can be difficult when different organisations use different benchmarks, definitions and testing methods.
A common standards framework could potentially make it easier to compare certain categories of AI-system behaviour.
For example, organisations could agree on testing methods for capabilities involving cybersecurity, autonomous operation, deception, dangerous biological information or other high-risk areas.
That does not necessarily mean every company would have to use exactly the same model or safety system.
Instead, a standards organisation could establish testing procedures that different companies could apply to their own models.
The Talks Are Still Developing
The current discussions should not be interpreted as confirmation that OpenAI, Anthropic and Google have already created a formal regulator.
Reports indicate that the companies have been discussing the idea, but important questions remain around how such a body would be structured and what powers it would have.
Those questions matter because an industry organisation can take several possible forms.
It could operate primarily as a technical standards group, publish recommended testing procedures or coordinate research.
A stronger structure could potentially have certification or enforcement mechanisms.
Exactly where the participants ultimately land remains uncertain.
AI Companies Are Under More Pressure
The timing of the talks reflects the rapid development of frontier AI.
Companies are releasing models that can reason, write software, interact with external tools, analyse information and operate for longer periods with less direct human instruction.
As those capabilities grow, the consequences of model failures can also become more significant.
A common testing framework could help developers identify certain problems before models are deployed widely.
However, standards alone would not eliminate AI risks.
Testing methods can have limitations, and model behaviour can change depending on deployment conditions, tools and users.
The quality of any future organisation would therefore depend heavily on its independence, technical expertise, transparency and willingness to update its methods as AI systems evolve.
What Happens Next
For now, the most important fact is that discussions are taking place.
There is no final structure to report, and the participating companies have not announced a fully established FINRA-style AI regulator.
If a formal body eventually emerges, its creation could become an important development in how the AI industry evaluates frontier models.
The discussion also demonstrates how the AI sector is beginning to move toward common infrastructure not only for computing and model development but also for safety evaluation.
As AI systems become increasingly capable, the debate is shifting from whether individual companies should test their models to how those tests can become more consistent, transparent and useful across the industry.


