Google, OpenAI and Anthropic reportedly plan their own body to set frontier AI's pre-release tests
Google, OpenAI and Anthropic are reportedly building SAFA, a government-independent body to set pre-release AI testing rules, aiming to launch in early 2027. Its members, evaluators and enforcement powers are unannounced, so buyers still have to set testing terms with each vendor themselves.
The Scientist · Science desk
What happened
- The Information reports that Google, OpenAI and Anthropic are working together on a body tentatively named the Standards Authority for Frontier AI, or SAFA.
- As described, SAFA would operate outside government control and write guidelines for risk assessment, testing and pre-release review of frontier models.
- People close to the plan say the aim is to launch the body officially in early 2027.
- The report came in the same week that Anthropic's Dario Amodei and OpenAI's Sam Altman urged the United Nations to create safeguards around AI.
Why it matters
- What it limits constraint Until someone is named to run the evaluations and a consequence is set for failing them, a SAFA result cannot replace a buyer's own due diligence on a model.
- Who has to decide decision Buyers who want pre-release assurance before SAFA exists have to write it into their vendor contracts themselves, because the body would arrive in late 2026 at the earliest.
- What it sets up precedent If SAFA launches as described, the first shared pre-release baseline for frontier models would be written by the developers it covers, so the charter's terms on who can join and who evaluates will decide how much weight it carries.
- What doesn't add up contradiction OpenAI also advocates mandatory US safety rules, so the idea that industry would set the rules in place of regulators matches only half of one founder's stated position.
Next story · Invest · Oct 1
Huang and Zuckerberg pushed back on Amodei's AI risk warnings at a closed White House meetingConfirmed6 publishers
A testing standard is only useful to the people downstream of it if two questions are settled: who runs the test, and what a failing result triggers. According to Superpower Daily, the companies have not announced membership, standards or enforcement powers, and the proposal assigns neither the evaluator's job nor the consequence of falling short [6][7]. A shared method would let buyers compare how models were tested, while leaving open whether anyone outside the developer checked the result [6]. Superpower Daily also gives a wider launch window than CIO, late 2026 or 2027 [4].
An industry venue already exists. All three companies belong to the Frontier Model Forum, founded in 2023 to advance safety research and best practices [5]. Superpower Daily asks whether SAFA would develop testing methods, assess whether members meet them, or simply publish guidance, and notes that those are different jobs [5]. A forum for swapping research need not hold authority over its members, the outlet adds. A standards body might do more, but its name cannot by itself establish what it would require of a developer whose model falls short [8].
OpenAI's blog post this week put industry standards next to government bodies. The company called for agreed-upon baselines, common measurements and incident reporting protocols, and wrote that such international standards "may be as important to pacing the frontier as alignment research itself" [9]. It named the US Center for AI Standards and Innovation, other federal frameworks, state regulations and new public-private partnerships as groups that can work together on that goal [10].
In the account of independent analyst Carmi Levy, buyers care less about global AI safety than about day-to-day exposure: agents leaking corporate or employee data, platform vulnerabilities, hallucinations in workflows, and compliance [12]. "They're worried about who's responsible for what, when (not if) AI goes off the rails," Levy said [13]. After a summer of reports of agents breaking containment, he said, AI companies are in no position to give that assurance [14]. He advised enforcing vendor standards now. One option is to require every new model to ship with the "equivalent of a safety and security datasheet" setting out its capabilities, known failure modes and testing history [15].
I think the reported design supports half of the idea that industry would set the rules buyers come to lean on. If SAFA launches as described, the labs would write the pre-release baseline for their own models, outside government control [2]. Whether buyers can rely on it depends on choices Superpower Daily lists as still open: who joins, how smaller developers are treated, and how independent any evaluators are [16].
What to watch
- Whether SAFA's charter names who runs pre-release evaluations and requires members to act when a model fails them.
- Whether SAFA's remit is formally separated from the Frontier Model Forum's, and whether smaller developers can join on equal terms.
- Whether US lawmakers take up the mandatory safety requirements OpenAI advocates, or CAISI gets a role in SAFA's testing.
Clarity's read
What the record supports and how the coverage leans. The claims behind it follow.
Reality
- Evidence35
- Adoption
- Not scored
- Hype gap+25
- Incentives65
- Confidence40
Claim ledger
Ranked by verification strength, evidence, and original report placement.
- [1]
According to The Information, Google, OpenAI and Anthropic are reportedly working together to create a body tentatively called the Standards Authority for Frontier AI (SAFA).
- [2]
SAFA would operate independently of government control and set guidelines around risk assessment, testing, and pre-release review practices for frontier AI models.
- [3]
Sources close to the matter say the goal is to officially launch the initiative in early 2027.
- [4]
The proposed group could launch in late 2026 or 2027.
- [5]
All three companies already belong to the Frontier Model Forum, founded in 2023 to advance safety research and best practices; the new body's distinct role has not been explained. Superpower Daily asks whether it will develop testing methods, assess whether members meet them or simply publish guidance, and says those are different jobs.
- [6]
A common test would still leave open who evaluates models, what results require action and whether participants must comply; none of those powers has been publicly specified and the proposal assigns neither the evaluator role nor the consequence of failure.
- [7]
No membership, standards or enforcement powers have been announced for the proposed body.
- [8]
A forum for exchanging research and best practices need not have authority over its participants. A standards body might do more, but its name alone cannot establish what it would require of a developer whose model falls short.
- [9]
In a blog post this week, OpenAI contended that safe and beneficial AI progress requires advances in alignment research along with shared standards including agreed-upon baselines, common measurements, and incident reporting protocols, and that such international standards "may be as important to pacing the frontier as alignment research itself."
- [10]
OpenAI said existing groups like the US Center for AI Standards and Innovation (CAISI), other federal frameworks, state regulations, and new forms of public-private partnerships can work together to advance this mission.
- [11]
OpenAI has said it supports working with other frontier labs on voluntary standards, with or without government support, and it also advocates mandatory U.S. safety requirements.
- [12]
Enterprises care less about the global state of AI safety than about how it will impact their day-to-day operational security, including whether an AI agent will expose corporate or employee data, platform vulnerabilities, hallucinations in corporate workflows, and AI compromising compliance efforts.
- [13]
"They're worried about who's responsible for what, when (not if) AI goes off the rails,"
- [14]
As the industry grapples with a summer's worth of reports of agents breaking containment, AI companies are in no position to provide these kinds of assurances.
- [15]
Levy advised enterprises to enforce standards for vendors now rather than waiting for something like SAFA, which could involve requiring every new model to ship with the "equivalent of a safety and security datasheet" outlining the model's capabilities, known failure modes, and testing history.
- [16]
The body's eventual membership, treatment of smaller developers and independence of any evaluators would help show whether it serves as a credible check or primarily reflects the priorities of its founders; those design choices remain open.
- [17]
The news came the same week that heads of leading AI companies, including Anthropic CEO Dario Amodei and OpenAI CEO Sam Altman, urged the United Nations to create safeguards around the technology they are building.
Sources
2 independent publishers whose own reporting we read for this story.
- cio.comThe companies racing to build frontier AI are now racing to govern it | CIO
1 article · September 24, 2026
- superpowerdaily.comGoogle, OpenAI and Anthropic Are Reportedly Working on an AI Safety Standards Body | Superpower Daily
1 article · September 24, 2026
Topics and entities
Entities
- OpenAI
- Center for AI Standards and Innovation (CAISI)
- The Information
- Standards Authority for Frontier AI
- Sam Altman
- Dario Amodei
- Frontier Model Forum
- Anthropic