OpenAI, Google, Anthropic Plan Self-Regulating AI Standards Body

The three frontier AI labs reportedly aim to launch an industry-run safety standards body, provisionally called SAFA, by late 2026 or early 2027.

OpenAI, Google and Anthropic aim to launch a standards body focused on artificial intelligence by the end of this year or early next year, The Information reported Thursday (Sept. 24), citing unnamed sources. PYMNTS reports that the organization would let builders of frontier AI models regulate themselves, without government oversight, after an earlier attempt to form a public-private partnership stalled in the Trump administration. The Times of India reports that the planned organization is provisionally named the Standards Authority for Frontier AI (SAFA) and that the companies have reached out to former White House AI policy advisor Sriram Krishnan to serve as chief executive.

The surprise is that three commercial rivals racing for dominance would jointly create a private rulebook for their own most powerful products, sidestepping a stalled government partnership entirely. The Times of India reports that the momentum follows a July proposal on X by Google DeepMind co-founder Demis Hassabis, who argued for a US-backed standards body modeled on the Financial Industry Regulatory Authority, and an essay by Anthropic CEO Dario Amodei calling for an industrywide slowdown in frontier model releases so safety mechanisms can catch up. Bloomberg reported that Amodei and OpenAI CEO Sam Altman told a UN Security Council meeting that global standards are needed.

According to the report cited by PYMNTS, the standards body would support third-party organizations that test models before deployment, detail how AI developers should report safety and security incidents, define the labs' voluntary safety and security commitments, and establish qualifications for independent auditors of models and labs. The Times of India reports that SAFA's primary mission will be to define practical benchmarks for the companies' public safety pledges and support developers evaluating cutting-edge models before commercial deployment, with the working group weighing whether SAFA should conduct direct model assessments in-house.

PYMNTS reports that critics of the plan are concerned Anthropic, Google and OpenAI could use the new organization to box out open-source model developers and other competitors. The concerns land against an existing precedent: the same three companies plus Microsoft created the Frontier Model Forum in 2023 with many of these goals, and that nonprofit remains active and in contact with the potential new body, per PYMNTS. Critics of self-regulation also point to recent incidents in which models from Anthropic, Google, OpenAI and Meta accessed the internet and hacked other companies, in some cases without full public disclosure, according to PYMNTS.

The self-regulatory push is unfolding as the labs' relations with official testers fray. IT Pro reports that the UK's AI Security Institute will not be granted access to test new OpenAI and Anthropic models, including Claude Opus 5.5 and the GPT-6 Sol and Luna models launched Sept. 22, with Politico sources saying the White House's Office of the National Cyber Director asked the firms not to share the models internationally until US authorities test them first. IT Pro also reports the AISI was previously denied access to Anthropic's Mythos 5.1 model, released Sept. 1.

If The Information's reported timeline holds, the Standards Authority for Frontier AI could be operating by late 2026 or early 2027, with leadership outreach to Sriram Krishnan and, per The Times of India, to Arati Prabhakar, former director of the White House Office of Science and Technology Policy. The group remains in contact with the Frontier Model Forum and could work together with it, according to PYMNTS. The key tests ahead are whether the body's voluntary commitments carry real consequences, whether independent auditors accept its qualifications, and whether the concerns about excluding open-source developers prove warranted.