Anthropic, OpenAI, and Google Talk Shared AI Safety Standards Body

Anthropic, OpenAI, and Google are discussing a shared AI safety standards body for testing and auditing frontier models, Asia Business Daily reported on 14 September 2026, citing mid-September summaries of The Information. Working-group meetings among the labs have reportedly run since July 2026, months after Google DeepMind CEO Demis Hassabis floated a U.S. Frontier AI Standards Body modeled on FINRA on 14 July. The talks sit beside Dario Amodei's public essay We Must Pace the Frontier, which urged independent evaluators with employee-like access, common standards among democratic labs, and international cooperation—including with China where possible—while comparing a standards organization to arms-control frameworks such as SALT.

What the private working group is exploring

The angle that matters for 14 September coverage is not another reprint of Amodei's essay. It is the claim that Anthropic, OpenAI, and Google have been meeting privately to weigh a joint standard-setting organization whose job would be testing and auditing safety—closer to a shared referee function than to a single lab's internal red team. Outlets including InvestingLive and Cryptopolitan amplified the same Information-sourced thread on 13–14 September, framing the discussion as industry-level standards rather than a one-off research collaboration.

Hassabis's July proposal supplies the institutional metaphor. FINRA is a self-regulatory body for U.S. broker-dealers; a Frontier AI Standards Body modeled on that idea would, in theory, set common evaluation expectations and audit practices for labs shipping the most capable systems. Whether such a body would be voluntary, government-chartered, or hybrid remains unspecified in the public summaries. What is clear from the reporting chain is that three of the highest-profile Western labs are treating standards infrastructure as a live agenda item, not only as op-ed rhetoric.

Sam Altman has publicly called independent evaluators with employee-equivalent access an excellent idea. Hassabis has said the direction is right and noted he recently proposed an industry-level standards organization. Microsoft CEO Satya Nadella has welcomed a measured pace for alignment work. Those statements do not prove a charter is drafted, but they do show senior executives converging on language about shared evaluation norms. Readers can set this private-talks story beside our earlier coverage of Amodei's pace-the-frontier essay and Google's internal shift of its AI responsibility team out of DeepMind.

Why standards talk surged after July

July 2026 concentrated several pressures. OpenAI agents briefly touched a Hugging Face incident that became a cautionary anecdote in subsequent policy coverage. Recursive self-improvement risk—systems that could accelerate their own capability gains—remains a stated concern among frontier researchers. Meanwhile U.S.–China competition continues to argue against any unilateral slowdown that only one geopolitical bloc observes. Korea University professor Choi Byungho, quoted in the Asia Business Daily thread, underscored that competitive dynamics may limit how far voluntary pacing can go even if labs agree on audit templates.

Amodei's essay sketched a package: independent evaluators who can see what employees see; common standards among democratic labs; and cooperation with China where feasible. He compared a standards organization to SALT-style arms-control architecture—an analogy that signals verification and mutual restraint more than marketing alliances. A private working group among Anthropic, OpenAI, and Google would be a narrower first step: agree how to test and audit before asking governments or rivals to join. Legislative and agency context for agent and model governance already appears in our Stop Rogue AI Act and NIST agent standards briefing and in the EU AI Office's first information requests.

The risk for critics is capture: if the same labs that ship frontier systems also design the scoring rubrics, audits may normalize industry comfort zones. The opportunity for supporters is interoperability: shared test harnesses, disclosure formats, and evaluator access rules could make third-party audits cheaper and more comparable across products. Neither outcome is settled by a working-group leak; both explain why journalists treat the talks as policy news rather than product rumor.

What editors should watch next

Near-term observables are concrete. Watch for a named entity—consortium, nonprofit, or government-sponsored body—with a charter that lists testing and auditing duties. Watch for whether employee-equivalent evaluator access becomes a written commitment or stays a speech-line. Watch for whether Chinese labs or governments are invited into any parallel track, matching Amodei's international-cooperation language, or whether the project remains a U.S.–aligned club. Watch also for friction with existing regulators: a FINRA-like body that duplicates EU AI Office or NIST work will face legitimacy questions from day one.

Documented facts for this draft stay narrow and attributable. On 14 September 2026 Asia Business Daily and peer outlets reported Anthropic, OpenAI, and Google discussing a joint AI safety standards organization after working-group meetings since July; Hassabis proposed a FINRA-modeled Frontier AI Standards Body on 14 July; Amodei, Altman, Hassabis, and Nadella have publicly endorsed independent evaluation and measured pacing language; and competitive and RSI concerns remain the cited constraints on voluntary slowdowns. Until a charter appears, the story is alignment of intent among major labs—not a finished regulator.

Primary threads are summarized via Asia Business Daily's English report, InvestingLive's standards-body brief, and Cryptopolitan's shared-standards coverage.

• Asia Business Daily: Anthropic, OpenAI, Google AI standards talks
• InvestingLive: Labs held talks on AI safety body
• Cryptopolitan: Shared AI standards body weighing
• PromptCrates: Amodei pace the frontier
• PromptCrates: Google moves AI responsibility from DeepMind
• PromptCrates: Stop Rogue AI Act and NIST agent standards
• PromptCrates: EU AI Office first information requests