Google, OpenAI and Anthropic want one safety rulebook for AI
AI safety talks among Google, OpenAI, and Anthropic could result in an independent body by late 2026 or early 2027, focused on model testing and practical safety benchmarks.
Three AI rivals may be working on the same safety checklist. Google, OpenAI and Anthropic are reportedly working on an independent body that could set common standards for testing advanced AI models before they are widely deployed.
The proposed organisation, provisionally called the Standards Authority for Frontier AI, or SAFA, could launch by the end of 2026 or early 2027, according to reporting by The Information. The group is expected to operate independently of government oversight.
The idea is to turn broad safety commitments made by AI companies into practical benchmarks, testing procedures and reporting standards that can be applied across frontier models.
From safety promises to measurable tests
AI companies already publish their own safety commitments, but the way those commitments are tested can vary from one lab to another. SAFA could provide a common framework for evaluating advanced models before commercial deployment.
The proposed body would support third-party testing organisations and establish standards for independent auditors, while discussions are also underway about whether SAFA itself should conduct some model evaluations.
The framework could cover areas such as AI safety and security testing, risk assessments and how developers report significant incidents. That could make it easier to compare how different frontier models perform against similar safety requirements, rather than leaving each company to define its own testing approach.
Why three competitors are joining forces
Google, OpenAI and Anthropic are competing directly in the AI market, making the proposed collaboration unusual. But the companies face some of the same questions as their models become more capable.
Governments and independent organisations need ways to assess these systems before they reach wider use, while AI developers need testing methods that can keep pace with increasingly complex models.
The proposal also comes as questions remain over whether existing government testing resources are sufficient for frontier AI. Reports say some people involved in the discussions believe the US Commerce Department's Center for AI Standards and Innovation may not have enough resources to evaluate the most advanced systems in depth.
Who could run SAFA?
The organisation's leadership has not been finalised. The companies have reportedly approached former White House AI policy adviser Sriram Krishnan and former Office of Science and Technology Policy director Arati Prabhakar about senior roles.
Other names reportedly considered for leadership or advisory positions include former US Secretary of State Condoleezza Rice, investor David Friedberg, METR founder Beth Barnes and AI researcher Paul Christiano.
For now, SAFA remains a proposal rather than an operating regulator. Its name, structure, leadership and exact powers are still under discussion.
That leaves the most important question for later: whether a safety framework created by the companies building frontier AI can earn enough independence and wider industry participation to become a standard others actually use.