OpenAI, Anthropic and Google confirm months-long informal safety consultations; talks underway since July

On September 15, 2026, OpenAI publicly confirmed that it has been holding informal safety consultations with rivals Anthropic and Google DeepMind for "several weeks," and its spokesperson told CNBC that the discussions have been ongoing since Google DeepMind chair Demis Hassabis released a July proposal for a U.S.-led AI "Standards Body" (CNBC, FACT). The Information first reported on Sep 13-14 that the three companies have held working-group meetings since July aimed at creating an industry-led AI safety standards body, citing people familiar with the discussions (INDEPENDENT EVIDENCE). CNN independently corroborated on Sep 14 with its own sources (INDEPENDENT EVIDENCE).

At a briefing in Washington the same day, OpenAI's chief global affairs officer Chris Lehane said: "It's better to try to work together to prioritize safety" (as reported by Bloomberg; FACT). He stated the engagement had been under way for several weeks, that OpenAI does not believe the three firms need an antitrust waiver to coordinate on safety, and that OpenAI would support bipartisan legislation aimed at mitigating catastrophic AI risks (FACT, as reported). Lehane also framed the coordination as precedent-based, citing the airline industry's safety coordination (Quartz; FACT as reported).

The confirmation lands inside a week of unusually public safety activity: Anthropic CEO Dario Amodei's Sep 12 essay urging the industry to pace the frontier (context, S15), endorsements from OpenAI's Sam Altman, xAI's Elon Musk and Hassabis, and FTC Chair Andrew Ferguson's same-day (Sep 15) remark that he is "deeply suspicious" of antitrust-exemption requests — "That sure sounds like moat digging" (Quartz citing Bloomberg; FACT).

Sequence of events:

• 2026-07-14 — Hassabis publishes X essay "A Framework for Frontier AI and the Dawning of a New Age" proposing a US-led FINRA-style Standards Body (verified directly; FACT).
• July 2026 — Informal tri-lateral working-group consultations begin (The Information/CNN; INDEPENDENT EVIDENCE).
• 2026-09-09 — Lehane publishes "The AI policy window is open. We need to act," committing OpenAI to building "industry-led standards" with other frontier labs "with or without government support" (OpenAI blog; FACT that it was published, content = COMPANY CLAIM).
• 2026-09-13/14 — The Information first reports the talks; CNN corroborates.
• 2026-09-15 (event date) — OpenAI confirms; Lehane briefs in Washington; Reuters wires Bloomberg's report; FTC Chair Ferguson voices antitrust scepticism.
• 2026-09-16/17 — Follow-through coverage; Anthropic and Google DeepMind still not confirming on the record (TyN: "leaving OpenAI as the only company to confirm the discussions").

• From speculation to confirmed channel: The existence of inter-lab safety coordination moved from an anonymous-sourced story (The Information, CNN) to an on-the-record confirmation by one of the three participants (OpenAI), with the other two neither confirming nor denying (INDEPENDENT EVIDENCE; COMPANY CLAIM only for OpenAI).
• Timeline established: The talks are now publicly dated to July 2026 — predating Amodei's Sep 12 pacing essay and the public resignation of a former Anthropic researcher, per OpenAI's spokesperson and The Information (FACT as reported).
• Anchor point named: OpenAI tied the start of the consultations to Hassabis's July 14 Standards Body proposal (CNBC; COMPANY CLAIM corroborated by the verified essay date).
• Antitrust framing opened: Lehane's assertion that no waiver is needed, and Ferguson's "moat digging" riposte, put the coordination squarely inside the U.S. antitrust debate for the first time (FACT as reported; INTERPRETATION below).
• Governance posture shift: A week of public essays (Amodei Sep 12, Lehane Sep 9) plus a confirmed private channel means the leading frontier labs are now visibly building a de-facto private governance layer beneath formal regulation — the story's core significance.

Before (pre-July 2026): The frontier labs' safety coordination was limited to formal, often performative venues: the 2023 Frontier Model Forum and AI Safety Fund, individual Responsible Scaling Policies, a June 2026 White House executive order on voluntary AI-cybersecurity safety reviews, and an Aug 4, 2026 White House meeting (Bloomberg, background). Public posture was competitive: no confirmed standing channel for discussing joint safety standards, pre-release review norms, or slowdown coordination.

Change (July → Sep 15-17, 2026): Hassabis's July 14 essay publicly proposed the Standards Body concept; behind the scenes, OpenAI, Anthropic and Google began working-group consultations (INDEPENDENT EVIDENCE). OpenAI confirmed the channel on Sep 15 and endorsed industry-led standards built "with or without government support" (COMPANY CLAIM). The three labs then publicly endorsed Amodei's pacing essay, converting private alignment into public unity (FACT as reported). Ferguson's FTC pushed back on the antitrust dimension the same day.

After: There is now a confirmed, months-old, informal tri-lateral safety-coordination channel among the three most powerful U.S. frontier labs, openly linked to a concrete proposal (FINRA-style Standards Body) that would — if built — perform pre-release evaluations, define "Frontier-class" thresholds, and potentially coordinate slowdowns. The channel exists independent of, and in parallel to, government frameworks (The Information/CNN: talks continue "with or without the support of the Trump administration").

As reported (no public charter, meeting minutes, or membership rules yet — this is an informal consultation, so "how it works" is inferred from reporting):

• Format: Working-group meetings at executive/policy level among the three labs, described by The Information as held "since July," meeting "as recently as last week" (as of Sep 13-14) (INDEPENDENT EVIDENCE).
• Catalyst document: Hassabis's July 14 essay functions as the shared reference model: a US-led Standards Body modelled on FINRA — a federally overseen public-private partnership or self-regulatory organization — with a board of independent technical experts and open-source representatives, funded largely by industry, running frontier-model benchmarks with voluntary 30-day pre-release reviews that could become mandatory once the protocol is proven, and with authority to coordinate industry-wide slowdowns if dangers escalate (Hassabis essay, verified directly; FACT that this is the proposal).
• Division of preference: Per The Information (INDEPENDENT EVIDENCE), the three companies differ on government involvement — Anthropic leaning toward government partnership, OpenAI emphasizing voluntary industry standards while also supporting state legislation, and Altman telling employees the labs may have to build the body themselves absent U.S. government backing.
• Public choreography: Private consultations are now synchronized with public signaling — essays (Lehane Sep 9, Amodei Sep 12), cross-endorsements (Hassabis, Altman, Musk), and OpenAI's Sep 15 confirmation — amplifying pressure on lawmakers (Quartz: OpenAI supports the Obernolte-Trahan House framework, the Thune-Cruz-Klobuchar Senate effort, and the FRONTIER Act's independent-verification provision).

This is the first confirmed multi-lab safety-coordination channel among OpenAI, Anthropic and Google — the three labs that effectively define the U.S. frontier. Its significance:

• A new de-facto governance layer below regulation: If competitors that control the frontier quietly coordinate safety standards, pre-release testing norms and slowdown decisions, they become a private regulator of last resort — precisely because Congress is gridlocked (Quartz: any AI bill becoming law before the end of 2026 is a long shot; the House departed for the November elections).
• Antitrust meets safety: The Lehane-Ferguson exchange frames the central tension: safety coordination among competitors can be read as legitimate joint risk management or as "moat digging" (Ferguson). How the U.S. resolves this will shape every future cross-lab agreement.
• Pacing debate institutionalized: The consultations give the "pace the frontier" argument (Amodei, S15; von der Leyen, S22) an actual mechanism — an industry channel through which a coordinated slowdown could in principle be negotiated.
• Precedent chain: It extends the Frontier Model Forum lineage but is qualitatively different — informal, CEO-adjacent, and aimed at enforceable-style standards rather than voluntary principles.

• Joint safety standards: Shared definitions of "Frontier-class" models, common evaluation protocols, and mutual recognition of pre-release testing — as outlined in Hassabis's essay.
• Coordinated deployment decisions: A channel through which "when to slow or stop" could be agreed across labs (Lehane's Sep 9 post explicitly contemplates "slowing the advancement of model capabilities").
• Fast-tracked legislation: The confirmed channel plus public unity gives lawmakers a concrete target for narrow antitrust carve-outs — reportedly already proposed by Sens. Banks and Schiff in a defense bill (Quartz citing Bloomberg) — and for the FRONTIER Act's independent-verification provisions (TechCrunch).
• A third-party evaluator ecosystem: Both the Hassabis framework and Altman's pledge (via TechCrunch) to embed third-party evaluators presume an independent-auditor industry — a new market segment.

1. Will Anthropic or Google ever confirm on the record — and if not, why the silence?
2. What is the actual agenda: joint evaluations, pre-release review, incident sharing, slowdown triggers, or all of the above?
3. Who convenes and chairs the working group? (The Information suggests Amodei is "driving the push"; Altman reportedly told staff labs must build it themselves.)
4. Will a concrete organizational form emerge (a new legal entity vs. a repurposed Frontier Model Forum vs. a FINRA-style SRO)?
5. Does the FTC/Ferguson escalate, and does the Banks-Schiff antitrust carve-out become law in the NDAA?
6. Will any of this bind actual release trajectories — e.g., did Sept 2026 frontier releases (GPT-6 Astra, Mythos 5.1) already move due to the channel?
7. How does the channel interact with the EU AI Act's Sep 15 GPAI evaluation deadline, the UK AISI (S03), and DeepMind's new institute (S04)?
8. Is there a role for international partners or open-weights labs, or is this a closed U.S.-only effort?

• Near term (weeks): Expect (a) the NDAA to carry or drop the Banks-Schiff antitrust carve-out — a strong tell on the channel's legal runway; (b) either an Anthropic/Google confirmation or continued studied silence; (c) consolidation with the DeepMind Institute (S04) and OpenAI's disclosure framework (S02) into a visible "standards package."
• Medium term (quarters): A concrete organizational form is likely — repurposed Frontier Model Forum or a new FINRA-style entity — with industry funding and independent-expert board seats per Hassabis's template. Expect the first annual/quarterly "frontier evaluation" pilot tied to a 2027 release.
• Political path: With Congress gridlocked until after November, the labs' "with or without government" stance tests whether private governance can precede law; post-election, the Obernolte-Trahan/Thune-Cruz-Klobuchar bills are the likeliest vehicles.
• Watch item: The first genuine stress test — a model release that the channel recommends delaying, and whether all three labs actually comply. Until then, the consultations remain a promise. PREDICTION (not FACT): some formalized standards body proposal from the three labs will surface before the end of 2026, conditioned on the antitrust carve-out.

The quiet confirmation that OpenAI, Anthropic and Google have been talking about safety since July is, on its face, one paragraph of press statements — but it is the load-bearing fact of the week's entire governance narrative. It converts "the industry should pace itself" from rhetoric into infrastructure: a real, months-old coordination channel exists, with a concrete FINRA-style design on the table and only one of three participants willing to say so publicly. The story's true drama is the double silence: Anthropic and Google won't confirm, and the FTC is already calling the whole project "moat digging." That is the shape of frontier AI governance in 2026 — private, informal, three companies wide, and already in an antitrust fight with the government it is trying not to need. Readers should track what the channel does (a withheld release, a shared evaluation, a carve-out in the NDAA), not what it says, because a coordination channel that produces nothing is just a press release with a longer meeting schedule.

Artifact: research/S24.md. Runtime window per RESEARCH_CONFIG.json: 2026-09-10 to 2026-09-17. Event date 2026-09-15 is in-window. Sources: sources/S24.md. Lab: labs/S24.md.