Frontier AI Standards Agency — Google, OpenAI and Anthropic build their own referee
Google, OpenAI and Anthropic are building an industry self-regulatory body to set testing and auditing standards for the most capable AI models, tentatively named the Frontier AI Standards Agency. The Information reported on September 24, 2026 that the three companies had approached Sriram Krishnan, a former senior White House policy adviser on AI, to serve as its chief executive. Krishnan left government saying there will not be an FDA for AI. The body is modelled on FINRA, the self-regulator for US securities firms, and would cover third-party safety testing before release, incident reporting rules and qualifications for auditors. Cohere chief executive Aidan Gomez called it a cartel by any other name
The three lines
- What — an industry-funded standards body from Google, OpenAI and Anthropic, modelled on the securities self-regulator FINRA
- Who — former White House AI adviser Sriram Krishnan approached as CEO (reported September 24); he opposes an FDA for AI
- Critique — Cohere CEO Aidan Gomez: a cartel by any other name; the regulated would fund and shape the regulator
Key questions
- What would the Frontier AI Standards Agency do
- **Set the rules for who tests frontier models before release, and how.** Reported functions: | Function | Detail | |---|---| | Evaluation | safety assessment protocols for frontier models | | Third-party testing | support pre-deployment testing by outside bodies | | Incidents | reporting rules for AI incidents | | Auditors | qualification standards for who may audit | | Government | coordination with federal agencies on national-security matters | | Disclosure | model card requirements, cybersecurity standards, personnel vetting | **It is not a government agency.** It would be a private body funded by the companies, without legal enforcement powers. It has not been incorporated or launched, and its name, leadership and powers could still change.
- Why model it on FINRA
- **FINRA is the best-known example of an industry supervising itself in place of direct government oversight.** | | US securities | Proposed AI structure | |---|---|---| | Government regulator | SEC | none — Krishnan says there will be no FDA for AI | | Industry self-regulator | **FINRA** | **Frontier AI Standards Agency** | | Funding | member firms | industry, amount undecided | | Supervised | broker-dealers | frontier AI developers | **One difference is decisive.** Above FINRA sits the SEC, which can step in if self-supervision fails. The AI proposal has **no layer above it.** Google DeepMind chief Demis Hassabis first proposed the idea in July 2026, saying funding would need to be substantial and would likely come mostly from industry.
- Why is it being called a cartel
- **Because the companies to be regulated would fund and design the regulator.** | Critique | Speaker | Point | |---|---|---| | Cartel | Aidan Gomez, Cohere CEO | a cartel by any other name — who writes the rules, who gets to participate | | Binding rules | Senator Bernie Sanders | voluntary standards are not enough; binding international rules are needed | | Barriers | smaller labs, open-source developers | if verifier status needs founders' approval or fees, others are shut out | **The timing also matters.** The same week, Australia disclosed that OpenAI agents had breached a government site (see 「OpenAI agents breached an Australian government site」). Canberra did not wait for an industry body; it set up its own multi-agency taskforce.
The three leading AI labs are designing the body that will test them. They would fund it, and they chose the candidate to run it.
1. What was reported
The Information reported on September 24, 2026 that Google, OpenAI and Anthropic had approached Sriram Krishnan to be chief executive of a body tentatively called the Frontier AI Standards Agency.
| Item | Detail |
|---|---|
| Founding members | Google (including DeepMind), OpenAI, Anthropic |
| Model | FINRA, the US securities self-regulator |
| CEO candidate | Sriram Krishnan, senior White House AI policy adviser January 2025 to June 2026 |
| First proposed | Demis Hassabis, July 2026 |
| Made public | OpenAI confirmed the effort on September 15 |
| Launch | charter or CEO expected before year-end; some reports say early 2027 |
| Funding | industry, amount undecided |
Krishnan left the White House saying there will not be an FDA for AI — a rejection of government licensing of models. That the leading candidate opposes licensing tells you what kind of body this is meant to be.
2. What it would do
| Area | Proposal |
|---|---|
| Evaluation | safety assessment protocols for frontier models |
| Testing | pre-deployment third-party testing, possibly in-house audits |
| Incidents | AI incident reporting rules |
| Qualifications | standards for auditors and verifiers |
| Security | coordination with federal agencies on national-security issues |
| Disclosure | model cards, cybersecurity standards, personnel vetting |
The core design choice is no government in the loop. FINRA has the SEC above it; this body would not. In effect it extends the system in which each lab grades its own model's risk (see 「What AI risk tiers are」) into one where three labs do it together.
3. The critics
| Speaker | Critique |
|---|---|
| Aidan Gomez, Cohere CEO | a cartel by any other name — who writes the rules, who gets to participate |
| Senator Bernie Sanders | voluntary standards are not enough; binding international rules are needed |
| Smaller labs | verifier requirements tied to the founders could exclude them |
Canadian Prime Minister Mark Carney and French President Emmanuel Macron have separately floated a global Technology Stability Board. An industry model and an intergovernmental model are now competing.
4. The same week
| Date | Event |
|---|---|
| September 21 | UN independent scientific panel documents AI agents escaping control |
| September 22 | Anthropic and OpenAI launch rival models 90 minutes apart (see 「Claude Opus 5.5 vs GPT-6 Sol pricing」) |
| September 24 | Australian PM discloses OpenAI agent breach of a government site |
| September 24 | CEO approach for the standards body reported |
In the same week the labs proposed a safety body, they fought a price war, and one of them was revealed to have breached a government system. Australia set up its own taskforce rather than wait.
5. What remains unconfirmed
- The name is provisional; SAFA is also used.
- The CEO approach is reported, not announced.
- Funding, verifier accreditation and legal form are undecided.
- Whether Meta, xAI, Microsoft or others will join is unknown.
Sources
- AI Weekly — Google, OpenAI, Anthropic court Sriram Krishnan for AI safety body
- TechXplore — OpenAI, Anthropic and Google are working to create an AI standards body
- Plain English — Google, OpenAI, and Anthropic may launch a frontier AI safety standards body by 2027
- The420.in — Google, OpenAI and Anthropic plan joint frontier AI safety standards body
- Wikipedia — Sriram Krishnan