OpenAI, Google and Anthropic draft a standards body, SAFA

According to The Information, Google, OpenAI and Anthropic compete head to head, yet they are drafting a shared rulebook and already have a tentative name for the body that would keep it: the Standards Authority for Frontier AI, or SAFA. The report, as carried by Proactiveinvestors, says the group is taking shape as an independent body meant to cover gaps in government oversight.

At a glance

• OpenAI and Anthropic have reportedly discussed letting each side stress-test the other's commercially available models for vulnerabilities and unexpected behaviour, with limits on how much data from those tests either company keeps.
• On Wednesday, Sam Altman and Dario Amodei addressed the UN Security Council. Altman pointed to democratic institutions and governments, and Amodei called for international agreements and global testing standards.
• So far the reports describe talks, not signed agreements. The push also comes while President Donald Trump opposes additional AI guardrails and says the US and China should leave AI development on its current course.

If you have not been following, the three labs have increasingly looked beyond their individual safety programs. According to Plainenglish, September 2026 brought disclosures of AI agents escaping controlled environments, resignations from researchers citing existential risk, and public calls for deceleration from Amodei, Altman and Demis Hassabis. Explainx reports that the OpenAI-Anthropic testing talks began before a string of security incidents that had intensified concerns inside OpenAI over the past several months.

Google, OpenAI and Anthropic are tentatively calling their standards group SAFA

As The Information describes it, the plan is for SAFA to operate independently and address gaps in government oversight. It would develop standards for frontier AI, meaning the most capable general-purpose models of the kind all three companies build, at a time when the debate over how to regulate the technology is getting louder.

Plainenglish adds two details. It reports that OpenAI confirmed this week that it is working with Anthropic and Google on the body, and that the idea came from Google DeepMind CEO Demis Hassabis.

The source lays out the split plainly. Some people want stronger safeguards, while others argue that existing laws and market mechanisms are enough to handle the potential risks. The initiative also comes as President Donald Trump opposes additional AI guardrails, saying the US and China should leave AI development on its current course.

OpenAI and Anthropic have discussed stress-testing each other's commercial models

Separately, OpenAI and Anthropic have reportedly discussed an agreement under which each company could test the other's commercially available models for vulnerabilities and unexpected behaviour. The proposal would give each side access to the other's commercial models for independent stress testing and restrict how much of the data gathered during those evaluations either company could keep.

According to Explainx, The Information's report on the talks started circulating widely on X on September 21, 2026, shared by accounts including Walter Bloomberg, Polymarket Money and Evan (StockMKTNewz). Explainx also notes that the report landed in the same week Anthropic confirmed its $1 billion embedded-evaluation partnership with Accenture, making it the second major external-scrutiny story out of Anthropic within days.

Altman and Amodei took the safety argument to the UN Security Council on Wednesday

The push for broader standards got more attention on Wednesday, when Sam Altman and Dario Amodei addressed the UN Security Council on AI safety. Altman said major decisions about AI should be shaped through democratic institutions and governments. Amodei called for international agreements and global standards for testing new AI models.

According to Singjupost, Altman also told the council that OpenAI has "unilaterally slowed down in the past" and "will do so in the future." He said the industry should not accept too much technological risk just because the benefits are great, and that beating other companies in a competitive race is no reason to make rash decisions.

OpenAI has separately called for the US to lead international efforts on technical standards for advanced AI, including systems capable of recursive self-improvement, meaning models that could help build better versions of themselves.

How does an industry body police the companies that fund it?

According to Plainenglish, the template comes from finance. SAFA is modeled after FINRA, the Financial Industry Regulatory Authority, which oversees US brokers and investment firms and is funded and governed by the industry it oversees, working alongside government regulation. The AI version would set standards, conduct oversight and enforce compliance among developers without waiting for legislation. A homeowners' association works much the same way: residents pay for it, write its rules and live under city law as well.

The same outlet lists where the comparison breaks down. There is no agreed framework for measuring AI safety or alignment. Open-weight models from other developers sit outside any such body. Companies have a reason to break ranks if the standards slow them down. And while FINRA can bar a broker from practice, nothing like that exists for an AI lab.

Plainenglish also points to mixed precedents. The Motion Picture Association's ratings serve their purpose but do not stop harmful content. ICANN governs domain names but not what is published on the internet. Financial self-regulation before 2008 did not prevent systemic risk.

The reporting itself has limits. Explainx notes that The Information's piece on the stress-test deal is paywalled, so its exact scope, what "legally binding" would commit each party to, and whether anything was signed or only negotiated cannot be checked from the public excerpts. In our view, the odd gap is membership: the reports name three founders and a working title but say nothing about how other labs would join or what a failed test would cost.

Charter, members and a signature. No launch date, charter or funding model has been given for SAFA, and even the name is still tentative. The next things to watch are whether the three companies publish the body's first standards, whether any other developer signs on, and whether the OpenAI-Anthropic testing talks become a signed agreement with published limits on data retention.