OpenAI, Anthropic and Google are reportedly working together on plans to establish a new organisation focused on the safety and evaluation of advanced artificial intelligence systems. The proposed organisation could be launched by the end of 2026 or in early 2027, according to reports citing people familiar with the discussions.
The organisation is provisionally being referred to as the Standards Authority for Frontier AI, or SAFA. It is expected to focus on developing practical standards for evaluating advanced AI models, reporting safety and security incidents, and supporting independent testing before powerful AI systems are released commercially. The proposed body would be industry-led rather than directly operated by a government agency.
The discussions come at a time when the capabilities of frontier AI models are expanding rapidly. Companies including OpenAI, Anthropic and Google are developing increasingly advanced systems that can generate content, write and analyse software, use external tools and perform multi-step tasks. At the same time, researchers, governments and technology companies are discussing how these systems should be evaluated for potential security, privacy and other risks.
According to The Information, which first reported details about the planned organisation, the three companies are considering SAFA as an independent standards body that could help establish common approaches to AI safety. The proposed organisation would not replace government regulation but could create common industry practices for testing and monitoring advanced models.
One of the proposed areas of work is the development of benchmarks for evaluating frontier AI models before their public or commercial deployment. Such benchmarks could provide a common way to assess whether an AI model meets particular safety requirements before it is made widely available.
The proposed body could also establish guidelines for reporting AI safety and security incidents. This could be important as AI systems become more capable of interacting with websites, software tools and other digital systems. A common reporting framework could allow companies and independent researchers to document serious incidents using consistent standards.
Another reported responsibility could involve supporting independent evaluators. The proposed organisation may establish qualifications or standards for third-party groups that assess advanced AI models. This could create a more consistent process for external testing across different AI companies.
There is also discussion about whether SAFA itself should directly conduct model evaluations. According to reports, members of the working group are considering whether the organisation should perform some safety and capability testing internally or primarily support independent organisations that carry out those assessments.
The proposal follows several weeks of discussions between OpenAI, Anthropic and Google about AI safety cooperation. OpenAI global policy chief Chris Lehane confirmed in September that the company had been working with Anthropic and Google DeepMind on AI safety issues for several weeks. The discussions were reported as an effort to coordinate approaches to risks associated with increasingly capable AI systems.
The idea also gained attention after Google DeepMind CEO and co-founder Demis Hassabis called for stronger coordination around AI safety standards. Reports said the proposal drew comparisons with industry self-regulatory organisations and public-private safety frameworks. The discussions have continued as technology companies face pressure to demonstrate how they intend to manage risks associated with frontier AI systems.
Leadership for the proposed organisation is also reportedly being discussed. Former White House AI policy adviser Sriram Krishnan and former White House Office of Science and Technology Policy director Arati Prabhakar have reportedly been considered for senior roles, including the possibility of leading the organisation. These discussions are not final appointments, and the structure of SAFA has not yet been formally announced.
The proposed organisation would operate in an environment where government agencies are already working on AI testing and standards. In the United States, the Center for AI Standards and Innovation has a role in federal AI evaluation efforts. Reports indicate that some people involved in the SAFA discussions believe additional independent technical capacity could be useful for evaluating increasingly complex frontier models.
An important aspect of the proposal is its reported independence from direct government administration. Earlier discussions reportedly considered a model involving greater government participation, but the current proposal is described as an industry-led organisation. The exact relationship between SAFA and government agencies remains subject to further discussion.
The potential creation of SAFA comes as AI companies face increasing questions about how advanced models should be tested before deployment. AI systems are increasingly capable of autonomous actions, software development, research assistance and interaction with external digital services. These capabilities can provide significant benefits but also create new challenges when systems behave unexpectedly or operate with access to external tools.
A common safety framework could potentially provide companies with shared terminology and procedures for evaluating these systems. However, the details of the proposed framework, including its membership, funding, governance structure and enforcement mechanisms, have not yet been publicly finalised.
The term SAFA should therefore currently be understood as the provisional name of a proposed organisation rather than an officially established international AI regulator. Reports indicate that discussions are continuing and that the target launch period is the end of 2026 or early 2027.
For users and businesses in India, the development could become relevant if the proposed standards influence how advanced AI models are tested and released internationally. However, there is currently no indication that SAFA would directly regulate AI systems in India or replace Indian laws and regulatory frameworks.
The initiative also reflects a broader shift in the AI industry toward cooperation on safety standards. OpenAI, Anthropic and Google remain competitors in the development and commercialisation of AI models, but their reported discussions indicate that they are also exploring common approaches to certain safety questions.
If launched, SAFA could become part of the wider global discussion over independent AI testing, model evaluations, incident reporting and pre-deployment safety checks. For now, the organisation remains a proposal under development, and its final name, structure, leadership and responsibilities could change before any formal launch.
