AI Safety Coordination: OpenAI's Surprising Antitrust Risk

AI safety coordination between three of America’s leading frontier labs is now on the record. On Tuesday 15 September 2026, Chris Lehane, OpenAI’s chief global affairs officer, told reporters in Washington that OpenAI had been working with Anthropic and Google DeepMind on safety for several weeks. Bloomberg first reported his remarks. Lehane also said the three companies did not need an antitrust waiver to coordinate, and summed up the logic in one line: “It’s better to try to work together to prioritize safety.”

The waiver remark matters most. Three days earlier, Anthropic chief executive Dario Amodei wrote that companies building artificial intelligence “can and should voluntarily work together to set standards”, but that the US government would need “to issue a narrow waiver for certain kinds of safety conversations”. OpenAI now says the talks can go ahead without one. Both statements can be true at the same time, and the reason turns on what the labs actually agree to do.

This article sets out what Lehane confirmed, traces the working group back to July, and compares the three legal routes open to rivals that want to cooperate. It also weighs the critics who call the effort a cartel. For background, our earlier coverage reads Amodei’s essay on pacing the frontier, follows the endorsement from Demis Hassabis, examines the Senate AI bill and reports on the separate US-China AI safety talks.

Table of contents

What Lehane Confirmed About AI Safety Coordination
How AI Safety Coordination Built Up From July to September
The Antitrust Problem Inside AI Safety Coordination
Three Legal Routes for AI Safety Coordination
The Standards Body Behind AI Safety Coordination
Where Each Lab Stands on AI Safety Coordination
Washington’s Part in AI Safety Coordination
The Case Against AI Safety Coordination
Who Is Missing From AI Safety Coordination
Signals That AI Safety Coordination Is Becoming Real
What AI Safety Coordination Means for Businesses Using AI
Frequently Asked Questions About AI Safety Coordination
References

What Lehane Confirmed About AI Safety Coordination

Lehane’s remarks were brief, and most of what is known comes from a handful of reports built on the same briefing. Bloomberg said OpenAI is “working on steps to address” safety issues with its “top competitors”, and described the move as an escalation of industry efforts to respond to concern that the technology poses an economic and security threat. Every later account of the AI safety coordination, from Reuters to TechCrunch, traces back to that report.

Talks “under way for several weeks”

Lehane said the company’s engagement with Anthropic and Google “had been under way for several weeks”, according to Bloomberg. He did not give a start date for the AI safety coordination, name the people involved or describe any document. Reuters, which carried the Bloomberg report, said it could not independently verify it, and none of the three companies immediately responded to its requests for comment. TechCrunch said it had also asked all three for comment.

No waiver needed for AI safety coordination

The most consequential line was legal rather than technical. Bloomberg reported that “OpenAI does not see the need for an antitrust waiver for the three AI firms to coordinate on safety matters.” A Stocktwits summary of the same report, published by Asianet Newsable, added that the discussions appeared to focus on cooperation on safety “rather than a broader agreement to slow or halt model development”. That distinction decides how much legal risk the AI safety coordination carries.

An aviation comparison

The Times of India reported that Lehane defended cooperation between rivals by comparing it to commercial aviation. “There are many instances over time where companies are trying to help each other on safety,” he said. The comparison points to the model of AI safety coordination OpenAI appears to have in mind: competitors that share safety lessons while still fighting over customers, prices and products.

A Washington audience for AI safety coordination

Lehane spoke in Washington, where he was meeting lawmakers. Reuters reported that he said OpenAI would support bipartisan legislation aimed at mitigating catastrophic AI risks. The Times of India named two packages: a House framework from Republican Jay Obernolte and Democrat Lori Trahan, and emerging Senate legislation led by Majority Leader John Thune, Republican Ted Cruz and Democrat Amy Klobuchar. The AI safety coordination news therefore arrived alongside a push for binding federal rules.

Claim	Who made it	Reported by	Status
Talks with Anthropic and Google DeepMind under way for several weeks	Chris Lehane, OpenAI	Bloomberg, Reuters, TechCrunch	Reuters could not verify; no company comment
No antitrust waiver needed to coordinate on safety	Lehane, per Bloomberg	Bloomberg, Reuters	OpenAI’s view only
“It’s better to try to work together to prioritize safety”	Lehane	Bloomberg, Reuters, Times of India	Direct quote
OpenAI supports the FRONTIER Act’s independent verification provision	Lehane	POLITICO, Washington Examiner, TechCrunch	Consistent with OpenAI’s 9 September letter
Talks focus on safety rather than a broader slowdown deal	Bloomberg reporting	Stocktwits via Asianet Newsable	Not addressed by the labs
OpenAI would back bipartisan catastrophic-risk legislation	Lehane	Reuters, Times of India	House and Senate packages named

How AI Safety Coordination Built Up From July to September

Lehane’s confirmation ended a nine-week sequence rather than arriving out of nowhere. The working group, the policy letters and the essays were public in pieces before Tuesday. Read in order, they show AI safety coordination forming step by step, with each document pointing at the next.

14 July: a blueprint from Hassabis

Demis Hassabis, then chief executive of Google DeepMind, proposed that the US establish “a new Standards Body modelled on a federally overseen public-private partnership or self-regulatory organisation, much like the Financial Industry Regulatory Authority (FINRA)”. Quartz, citing The Information, described that essay as the catalyst for AI safety coordination between the three companies.

From July: a working group meets

The Information reported on 13 September that representatives from Anthropic, OpenAI and Google had been meeting as a working group since at least July. Quartz relayed that the group met as recently as the previous week, while CTech put the latest meeting about a week and a half before its 15 September article. The two accounts differ only slightly, but they show how little of the calendar behind AI safety coordination is public.

23 July: two bills on the same day

The FRONTIER Act, H.R. 9925, was introduced in the House on 23 July. The same day, lawmakers introduced the bipartisan Collaboration on Adversarial Threats and Security Risks Act, led in the Senate by Jim Banks and Adam Schiff and in the House by Bob Latta and George Whitesides. The first would require independent verification of the largest developers. The second would give AI companies a narrow antitrust exemption for certain security coordination.

9 September: OpenAI’s policy letter

Lehane published “The AI policy window is open. We need to act.” on OpenAI’s website. It said OpenAI “will work with other frontier labs to advance frontier AI standards, building a voluntary effort now, with or without government support.” It also backed four California bills and called for mandatory federal safety rules. In hindsight, the letter announced the AI safety coordination six days before Lehane confirmed it.

12 September: Altman hints and Amodei publishes

Sam Altman was asked by Fortune why he did not sit down with Amodei, Elon Musk and Hassabis to plan for safety. “I think that will happen,” he said in the interview, published on 12 September, adding: “I’m not going to pre-announce private discussions that I think should be at some point shared as a group.” The same day, Amodei published “We Must Pace the Frontier”, a three-step plan built on embedded evaluators and industry coordination.

13 to 15 September: from leak to confirmation

The Information’s report landed on Sunday 13 September, as did a critical essay from Cohere. That day, House Speaker Mike Johnson told CNN there was “no consensus among them” on standards. Two days later, Lehane confirmed the talks on the record, turning a reported working group into acknowledged AI safety coordination.

Date	Event	Source
14 Jul 2026	Hassabis proposes a FINRA-style Standards Body for frontier AI	Hassabis essay; TechCrunch
From July	Anthropic, OpenAI and Google working group starts meeting	The Information, via Quartz
23 Jul 2026	FRONTIER Act and Banks–Schiff exemption bill introduced	GovTrack; House sponsors
9 Sep 2026	OpenAI pledges voluntary frontier standards “with or without government support”	OpenAI
12 Sep 2026	Altman tells Fortune a joint plan “will happen”; Amodei publishes his essay	Fortune; Amodei
13 Sep 2026	Working group reported; Cohere essay; Johnson on CNN	The Information; Cohere; Quartz
15 Sep 2026	Lehane confirms weeks of talks and says no waiver is needed	Bloomberg; Reuters; TechCrunch

The talks trace back more than two months before OpenAI confirmed them.

Days before Lehane’s 15 September confirmation

Hassabis framework, 14 July 63 days

FRONTIER Act and Banks–Schiff bills, 23 July 54 days

OpenAI policy letter, 9 September 6 days

Amodei essay, 12 September 3 days

Working group report and Cohere essay, 13 September 2 days

The Antitrust Problem Inside AI Safety Coordination

Every serious account of the AI safety coordination talks mentions antitrust law, because the three companies are direct competitors in a fast-growing market. The legal question is not whether rivals may speak to one another. It is what they agree to once they do, and whether any agreement limits the competition that customers would otherwise enjoy.

Why rivals talking worries enforcers

Section 1 of the Sherman Act declares illegal “every contract, combination in the form of trust or otherwise, or conspiracy, in restraint of trade or commerce”. In practice, US courts judge most collaborations between competitors under a “rule of reason”, weighing likely harms against likely benefits. Agreements that fix prices, divide markets or restrict output are treated far more harshly. A deal to delay model releases could be characterised as a restriction on output, which is why AI safety coordination draws legal attention.

Altman flagged the risk

TechCrunch reported that some executives, including Altman, have noted that such talks could put the companies at risk of violating antitrust law “if the coordination is found to suppress competition”. Fortune wrote that AI executives had previously suggested any coordinated slowdown “would need government consent to avoid falling foul of U.S. antitrust laws”. The concern about AI safety coordination is not only an outsider’s objection; it comes from inside the labs.

Amodei’s request for a narrow waiver

Amodei’s essay argues that AI companies “can and should voluntarily work together to set standards”. It then adds a legal condition: “For antitrust reasons, it’s helpful for the US government to mediate or at least enable these discussions — they don’t need to participate, but do need to issue a narrow waiver for certain kinds of safety conversations.” A footnote to the essay’s industry step repeats the point: “With government mediation or waivers of antitrust restrictions.”

Lehane’s view that talks need none

Lehane’s position on AI safety coordination is narrower than it first sounds. He did not say a pacing agreement would be lawful without government help. Bloomberg’s account is that the firms need no waiver “to coordinate on safety matters”. Sharing evaluation methods, incident definitions and testing ideas resembles the everyday standards work that competitors in many industries do. AI safety coordination of that kind can plausibly proceed under existing law.

Where the two positions meet

Amodei’s essay draws the same line. Describing the industry step, it says: “Some forms of coordination that would be impactful for pacing are legally challenging, and will require government support.” Lehane is talking about the first kind of cooperation, and Amodei is looking ahead to the second. So AI safety coordination can continue now, while its most ambitious goal, an agreed slowdown, stays legally unsettled.

The guidance that disappeared

Companies once had a federal rulebook for this. In April 2000, the FTC and the Justice Department issued Antitrust Guidelines for Collaborations Among Competitors, including safety zones for some joint ventures. On 11 December 2024 the agencies withdrew them after a 3–2 FTC vote, saying they “no longer provide reliable guidance”. Businesses were told to “review the relevant statutes and caselaw” instead. On 23 February 2026 the agencies asked for comment on new guidance, including on information and data sharing, with a 24 April deadline.

Voice	Position on an antitrust waiver	Scope they have in mind
Dario Amodei, Anthropic	Government must “issue a narrow waiver for certain kinds of safety conversations”	Industry standards and, with government support, pacing
Chris Lehane, OpenAI	No waiver needed “to coordinate on safety matters”	Safety cooperation, not a slowdown deal
Banks–Schiff bill	Narrow statutory exemption with advance written notice	Distillation, model theft, loss of control, cyber and CBRN threats
Aidan Gomez, Cohere	Opposed: “a cartel by any other name”	Wants open, evidence-based standards instead
Mike Johnson, House Speaker	“There’s no consensus among them”	A partnership of industry and lawmakers

Three Legal Routes for AI Safety Coordination

US law offers rivals more than one way to cooperate, and each fits a different version of the talks. None was designed for frontier AI, and none gives three dominant labs a free pass. Setting them side by side shows where AI safety coordination stands on firm ground and where it would need new law.

Route one: talks under the rule of reason

The simplest route needs no filing at all. Competitors can discuss safety practice, compare evaluation methods and publish shared recommendations, accepting that any challenge would be judged under the rule of reason. This appears to be the kind of AI safety coordination Lehane means. Its weakness is uncertainty: the 2000 guidelines are gone, and the agencies’ February 2026 consultation is the latest public step we found, so companies lean on statutes, case law and their own lawyers.

Route two: a registered standards body

The Standards Development Organization Advancement Act, signed on 22 June 2004, extended earlier antitrust protections to standards bodies. An organisation that notifies the FTC and the Justice Department gets rule-of-reason treatment for its standards work, special rules on attorneys’ fees, and liability limited to actual rather than treble damages. The agencies then publish a notice in the Federal Register. A standards body built out of today’s AI safety coordination could, in principle, use this route.

What the 2004 law demands

The statute defines a standards development organisation as one using procedures that incorporate “openness, balance of interests, due process, an appeals process, and consensus”. Protection does not extend to exchanges of information on “cost, sales, profitability, prices, marketing, or distribution” that are not reasonably required, or to agreements that would “allocate a market” or “set or restrain prices”. A private group of three rivals would struggle to meet the openness test without widening its membership.

Route three: the Banks–Schiff exemption

The Collaboration on Adversarial Threats and Security Risks Act would create a narrow antitrust exemption so AI companies can share information and coordinate on threats. Its sponsors list distillation attacks, model theft, loss of control, and cyber, chemical, biological, radiological and nuclear risks. Coordination that delays or limits deployment of high-risk models would require written notice in advance, and price fixing and market allocation would stay prohibited. Bloomberg, citing a person familiar with the matter, reported that the measure has been folded into the annual defence bill.

Why a closed group fits none of them neatly

Each route has a catch. Informal talks carry uncertainty, a registered standards body must open its doors, and the Banks–Schiff exemption is not yet law. That explains why Lehane stressed safety cooperation, why Amodei asked Washington for help, and why critics focus on who sits in the room. AI safety coordination that stays small is legally simplest only while it avoids pacing.

Route	Legal basis	Filing	Main limit	Fit for pacing
Informal safety talks	Sherman Act, judged under the rule of reason	None	No federal guidelines since December 2024	Weak: output limits draw scrutiny
Registered standards body	Standards Development Organization Advancement Act of 2004	Notice to FTC and DOJ	Openness, balance, due process, appeals and consensus required	Covers standards work, not prices or market allocation
Banks–Schiff exemption	Proposed statute	Advance written notice	Not yet law; limited list of threats	Written for delaying or limiting high-risk deployments

The main legal shelter for standards bodies is more than two decades old, while the AI-specific exemption is only weeks old and not yet law.

Years since each tool appeared, as of 15 September 2026

Standards Development Organization Advancement Act, signed 22 June 2004 22.2 years

Frontier Model Forum, launched 26 July 2023 3.1 years

Collaboration guidelines withdrawn, 11 December 2024 1.8 years

Banks–Schiff exemption bill, introduced 23 July 2026 0.15 years

The Standards Body Behind AI Safety Coordination

The AI safety coordination talks are not only about sharing notes. Quartz and The Information describe the goal as a shared industry standards body, and the design most often cited is the one Hassabis published in July. Its details explain both the appeal of AI safety coordination and the unease it provokes.

A body “much like” FINRA

Hassabis proposed a board that includes “independent leading technical experts and open-source representatives”. Funding, he wrote, “would need to be substantial and likely mostly come from industry, in order to attract world-class technical talent and provide the necessary compute resources for large-scale testing”. FINRA, his model, describes itself as a private not-for-profit self-regulatory organisation, supervised by the Securities and Exchange Commission and funded by member fees rather than taxpayers.

Thirty days, then a requirement

Under the framework, “Frontier Labs would voluntarily share models with the Standards Body for review up to 30 days before release.” Once the assessment protocol proved “effective and robust”, Hassabis wrote, “formalisation could quickly follow, meaning that Frontier Models would be required to pass it to be deployed in the US market.” Evaluations “would be regularly updated, perhaps quarterly to start”. For AI safety coordination, that path from voluntary to mandatory is the crucial design choice.

A slowdown clause and a startup exemption

Two lines anticipate the current debate. The framework “could be ratcheted up if the seriousness of the situation demands, including coordinating a slowdown in development among the Frontier Labs if deemed necessary”. It also says “any non-frontier models, say from startups or academia, would be exempt from this process”. The first line is the part that raises antitrust questions. The second is an answer, of sorts, to critics who fear a moat.

What the working group has said

Very little about AI safety coordination has come from the group itself. Quartz reported that Altman told an OpenAI town hall he favoured a testing and auditing body, but thought the major labs would need to build it independently of government support. He acknowledged that industry-wide agreement remained elusive and that certain companies were resistant. OpenAI chief scientist Jakub Pachocki said the company had been talking “to some external organizations about potential concrete standards we could put in place”, CNN reported.

An existing forum for AI safety coordination

The three labs already share one body. Anthropic, Google, Microsoft and OpenAI launched the Frontier Model Forum on 26 July 2023, and Meta and Amazon joined in May 2024. Its stated purposes include identifying best practices, supporting “standards development for frontier AI safety and security”, and facilitating information sharing. The proposed standards body would go further, reviewing models before release and potentially becoming mandatory.

Feature	Frontier Model Forum	Proposed Standards Body	FINRA
Status	Operating since July 2023	Proposed July 2026; under discussion	Operating self-regulator
Members	Anthropic, Google, Microsoft, OpenAI, Meta, Amazon	“Frontier Labs” meeting benchmark criteria	US brokerage firms
Oversight	Operating board of member representatives	Federally overseen partnership or self-regulator	Securities and Exchange Commission
Funding	Not detailed on its About page	“Mostly come from industry”	Member fees
Core task	Best practices, standards support, information sharing	Review of frontier models before release	Investor protection and market integrity
Binding?	No	Voluntary first, later required for the US market	Yes, for member firms

Where Each Lab Stands on AI Safety Coordination

The three participants arrive with different public commitments. Reading their own documents, rather than summaries, shows that AI safety coordination rests on overlapping but not identical positions, and that the gaps between them are as informative as the overlap.

OpenAI: standards “with or without government support”

OpenAI’s 9 September letter is the most explicit statement on AI safety coordination. Beyond the voluntary effort, it says “Any industry-led standards would complement—not replace—mandatory federal safeguards and democratic oversight.” It argues that “frontier laboratories largely set their own rules for managing frontier risks” today, and that independent verification would replace “that fragmented system of private governance”. Altman has also said OpenAI would follow Anthropic in embedding third-party evaluators.

Anthropic: evaluators first, then a waiver

Amodei’s essay commits Anthropic unilaterally to one step: embedded third-party evaluators with “employee-like access” to verify safety practices and report incidents. Coordination among companies is the second step, and the essay treats government help on antitrust as a precondition for its strongest forms. Anthropic is therefore the participant whose version of AI safety coordination depends most on Washington.

Google DeepMind: the author of the blueprint

Google’s contribution to AI safety coordination is the design itself. Hassabis, who became Google DeepMind’s chair and Alphabet’s chief scientist after stepping down as chief executive in August, endorsed Amodei’s essay on 12 September. “The details need working through, but the direction is correct for meeting this critical moment,” he wrote, adding that this was “why we recently put out our proposal for an industry-wide standards body for frontier AI.”

What none of them has published

None of the reports reviewed for this article describes a charter, a membership list, a funding model, a governance structure or a test protocol for the planned body. Google, OpenAI and Anthropic did not comment to CNN, according to Quartz. Until one of those documents appears, AI safety coordination remains a set of statements and meetings rather than an institution.

Lab	Public commitment	Role it wants for government	Key document
OpenAI	Voluntary frontier standards now; embedded evaluators; FRONTIER Act verification	Mandatory federal rules that industry standards complement	9 September letter; 15 September briefing
Anthropic	Embedded third-party evaluators, committed unilaterally	A narrow antitrust waiver and support for pacing	12 September essay
Google DeepMind	Standards body blueprint; endorsed Amodei’s direction	Federal oversight of a public-private body	14 July framework; 12 September post

Washington's Part in AI Safety Coordination

Lehane took the AI safety coordination message to Capitol Hill, and the timing matters. Congress is weighing several frontier AI bills, while the White House is openly hostile to safety-led regulation. The talks will be judged partly by what lawmakers do with them.

OpenAI backs the FRONTIER Act’s verification rules

The Frontier Risk Oversight, National Transparency, Independent Evaluation, and Reporting Act was introduced by Obernolte with Trahan and four other co-sponsors. According to POLITICO, as relayed by the Washington Examiner, Lehane said he spoke to lawmakers about “how we think about IVO and made clear that we can support that”. IVOs are independent verification organisations, which the bill would license to assess the largest developers.

What the bill actually requires

The bill text defines a frontier model as one trained using more than 10^26 integer or floating-point operations, and sets two tiers of developer by revenue and AI spending. The Commerce Secretary could issue emergency orders “suspending or restricting” a model that presents “an imminent catastrophic risk”. Violating such an order would carry a civil penalty of up to $10 million per violation, with each day counting separately. The bill would also preempt certain state laws.

Provision in H.R. 9925	What the introduced text says
Frontier model	Trained using more than 10^26 integer or floating-point operations
Large frontier developer	Revenue above $50 million and at least $1 billion of AI development spending over 36 months
Very large frontier developer	Revenue above $5 billion and at least $10 billion of AI development spending over 36 months
Emergency orders	Commerce may suspend or restrict a model presenting imminent catastrophic risk
Penalty for breaching an order	Up to $10 million per violation, each day a separate violation
New official	Under Secretary of Commerce for AI Security
States	Bars new state obligations on developers in covered subject areas

The Senate track

A separate Senate effort led by Thune, Cruz and Klobuchar would focus on catastrophic risks. Our Senate AI bill coverage compares conflicting accounts of how much power it would give Commerce to stop a release. Bloomberg reported that OpenAI is open to that proposal too, according to the Stocktwits summary.

A White House that calls it a hoax

President Donald Trump has dismissed safety warnings as a “hoax”. On Monday 14 September he wrote on Truth Social that “AI, and Data Centers, will be the Greatest Economic Development Engine in History” and that “It will not be stopped by brilliantly run Destructive Forces”. TechCrunch reported that David Sacks, his key AI adviser, called fears of existential risk overblown. Quartz reported that a White House draft executive order to create an AI regulatory body was effectively shelved before summer ended.

Congress sees no consensus yet

House Speaker Mike Johnson told CNN on 13 September that there is little agreement among the companies on what standards and guardrails should look like. “There’s no consensus among them,” he said. “So this has to be a partnership with the industry itself, with the corporations that are doing this and with the policy and lawmakers.” For AI safety coordination, that is both a warning and an invitation.

The Case Against AI Safety Coordination

The loudest objection to AI safety coordination comes from a rival that is not in the room. Critics accept that AI needs rules; they dispute who should write them. Their arguments deserve the same scrutiny as the labs’ claims, because they identify the conditions under which AI safety coordination could harm competition.

“A cartel by any other name”

Cohere chief executive Aidan Gomez published “Who Gets to Define the Rules for AI?” on 13 September. “AI needs guardrails. That is not the dispute and never has been,” he wrote. “The dispute is over who writes them, who gets to participate and whose interests the rules are protecting.” He accused “these oligopolies” of “requesting to bend competition rules and be permitted to dictate the terms for everyone else”, and concluded: “A wolf in sheep’s clothing, a cartel by any other name.”

A room with the door shut

Gomez framed the choice over AI safety coordination as procedural. The rules will be written either way, he argued; what remains open is “whether they get written by a group anyone can join and with evidence anyone can check, or by a handful of companies in a room with the door shut.” That argument echoes the openness requirement in the 2004 standards law, one reason a closed working group would struggle to claim its protections.

Lessons from earlier exemptions

Gomez also cited history. He pointed to credit rating agencies, a market with “only three of these evaluators” that went on to rate “subprime mortgage securities triple-A”. He cited Europe’s 1985 Motor Vehicle Block Exemption, won by car makers who argued that vehicles were complex, safety-critical machines. His argument is that exemptions justified by safety have entrenched incumbents before, and that AI safety coordination risks repeating the pattern.

Regulate the deployment, not the company

Cohere’s alternative is to set rules by use. Gomez wants standards applied “according to where a system is deployed and what it can touch, rather than how large the company that built it is”. He warned that “a small, poorly specified model sitting inside a hospital is a live risk today, and under a frontier-only regime nobody is even looking at it.” That cuts against both the FRONTIER Act’s size thresholds and the startup exemption Hassabis proposed.

Huang rejects the premise

Nvidia chief executive Jensen Huang attacked the warnings rather than the talks. Speaking at a conference in Los Angeles on 14 September, he said, according to AFP: “The confluence of these words and then the prediction is alarming and troubling and it shouldn’t be done. It’s irresponsible.” He added that the idea that AI could destroy humanity “is not grounded on science”.

A moat for the biggest labs

CTech’s analysis argued that safety rules could raise barriers for smaller competitors, because compliance costs weigh far more heavily on startups than on the largest incumbents. Meta chief executive Mark Zuckerberg also expressed concerns about plans for a national AI regulator in an August call with Trump, according to The Information as relayed by Quartz. OpenAI’s letter anticipates the charge: “A serious public framework should reduce, not increase, the concentration of power.”

Critic	Objection	Response on the record
Aidan Gomez, Cohere	Dominant labs want to bend competition rules	OpenAI: a framework should “reduce, not increase, the concentration of power”
Aidan Gomez, Cohere	Frontier-only rules miss risky small deployments	Hassabis exempts non-frontier models, which eases costs but not this risk
Jensen Huang, Nvidia	Doom predictions are “not grounded on science”	Altman: capabilities cannot go much further without progress on monitorability and alignment
Mark Zuckerberg, Meta	Concerns about a national AI regulator	Hassabis proposes an industry-funded, federally overseen body
Donald Trump	Safety warnings are a “hoax”	Johnson: rules need a partnership of industry and lawmakers
CTech analysis	Compliance costs build a moat	FRONTIER Act duties start only above its revenue and spending thresholds

Who Is Missing From AI Safety Coordination

Three labs are not the whole frontier. The companies outside the working group matter, because a slowdown agreed by some leaders and ignored by others would shift market share rather than reduce risk. That gap shapes how far AI safety coordination can go.

SpaceXAI and Elon Musk

Musk endorsed Amodei’s essay, as our coverage of Musk backing Amodei describes, yet no report places SpaceXAI in the working group. At the All-In Summit in Los Angeles on 14 September, he suggested a different mechanism, according to International Business Times: letting competitors, including “three or four of the leading Chinese companies”, run a test harness on each other’s systems.

Peer review instead of a standards body

Musk’s version would replace a central body with rivals checking rivals. “Instead of grading your own homework, you would at least have competitors grading your homework and raising the alarm if they see concerns,” he said. That idea overlaps with AI safety coordination but does not require the three current participants to agree on a single institution.

Meta, Microsoft and Amazon

Meta and Amazon joined the Frontier Model Forum in 2024, and Microsoft was a founder, yet none is named in reports of the standards body talks. Zuckerberg’s reported objection to a national regulator suggests Meta would not simply sign up. Microsoft and Amazon are long-standing investors in OpenAI and Anthropic respectively, which gives them influence over AI safety coordination without a seat at the table.

Chinese developers

Amodei’s essay treats China as the hardest problem. It sets out levels of global agreement, including one in which both sides test their models before release for acute risks, possibly through a global standards body. He wrote that creating such a body “is likely feasible, but giving it real teeth will be a challenge”. Our coverage of US-China AI safety talks tracks the separate government channel.

Startups and academia

Hassabis would exempt non-frontier developers, and the FRONTIER Act’s obligations start at more than $50 million in revenue and at least $1 billion of AI development spending over 36 months. Smaller players are therefore outside both the talks and most of the proposed duties. Gomez fears exactly that outcome, arguing that it leaves real risks in smaller AI models unexamined.

Signals That AI Safety Coordination Is Becoming Real

Statements are cheap; institutions leave paperwork. The following markers would show that AI safety coordination has moved from meetings to something enforceable, and each would appear somewhere public that readers can check.

A charter with open membership

A published charter for AI safety coordination would answer the questions critics are asking: who can join, who funds the body, who sits on its board and how decisions are appealed. Openness is also the test the 2004 standards law applies, so a charter would reveal whether the group intends to seek that protection.

A notice in the Federal Register

If the group chose to organise as a standards development organisation, it would need to notify the FTC and the Justice Department to limit its damages exposure, and the agencies would publish a notice. That filing would be the clearest sign of formal AI safety coordination.

The defence bill’s final text

Bloomberg reported that the Banks–Schiff exemption has been folded into the annual defence bill. Whether it survives, and in what form, will decide whether agreements to delay or limit risky deployments get a clear legal pathway.

Evaluators with badges

Amodei said Anthropic intends to invite an embedded external review team equipped with desks, access badges and company laptops, and Altman said OpenAI would do the same. The first published findings from those evaluators would give AI safety coordination something verifiable to coordinate around.

A first shared test

The strongest signal would be one model reviewed under a shared protocol before release, with results published. The 30-day window in the Hassabis framework offers a benchmark for how that might look in practice.

Signal	What it would show	Where to look
Published charter	Membership, funding, governance and appeals	Company websites and policy blogs
Standards body notification	A formal body seeking damages protection	Federal Register
Defence bill text	Whether the Banks–Schiff exemption survives	Congressional conference report
Embedded evaluator findings	Independent verification in practice	Anthropic, OpenAI and evaluator publications
Shared pre-release review	A working test protocol	Model cards and release notes

What AI Safety Coordination Means for Businesses Using AI

Most organisations will never sit in these talks, but they buy the products the talks govern. AI safety coordination among the largest labs could change contracts, release calendars and the questions procurement teams ask of their suppliers.

A common testing vocabulary

Channel Insider argued that a shared testing framework could let resellers and integrators compare competing AI products more consistently, giving buyers standard evaluation criteria similar to security certifications. If the labs converge on shared evaluations, expect vendor documentation to cite them. Organisations with an IT governance process can map those results into existing risk registers.

Release dates that move

A pre-release review window, voluntary or mandated, adds time between a model’s completion and its availability. Firms planning projects around specific model versions should build slack into roadmaps and keep fallbacks. Our analysis of AI safety measures makes the same point about the wider slowdown debate.

Questions to ask vendors now

Ask whether a provider gives independent evaluators access, which incidents it has disclosed, and which external standards its systems are tested against. Pair those answers with your own cybersecurity controls, reviewed with a security partner, and follow vendor changes through our AI models and tools hub. Treat AI safety coordination as a reason to ask sharper questions, not as a guarantee.

Frequently Asked Questions About AI Safety Coordination

Have OpenAI, Anthropic and Google signed a safety agreement?

No AI safety coordination agreement has been published. OpenAI confirmed several weeks of talks with Anthropic and Google DeepMind on 15 September, and The Information reported a working group meeting since at least July. House Speaker Mike Johnson said there was no consensus among the companies on standards.

Do the labs need an antitrust waiver for AI safety coordination?

OpenAI says not for coordination on safety matters. Amodei argues a narrow waiver is needed for certain safety conversations, and says some coordination on pacing is legally challenging without government support. The Banks–Schiff bill would create a limited exemption.

What would the proposed AI standards body do?

Based on the Hassabis framework, it would review frontier models up to 30 days before release, voluntarily at first, with passing later required for the US market. It would be funded mostly by industry, with independent experts and open-source representatives on its board.

Why does Cohere call the effort a cartel?

Aidan Gomez argues that a handful of dominant companies are seeking to bend competition rules and set terms for everyone else. He wants standards built through open participation and applied according to deployment risk rather than company size.

Does the White House support AI safety coordination?

It has given no public support. President Trump has called AI risk warnings a hoax, and Quartz reported that a draft executive order to create an AI regulator was shelved. CNN reported that the talks are proceeding with or without the administration’s backing.