OpenAI has confirmed it is in active talks with Anthropic and Google DeepMind to coordinate on AI safety, marking one of the most direct admissions yet that the industry’s fiercest rivals are quietly building a shared framework to manage risk from frontier models. The confirmation landed on September 15, 2026, and was reported within hours by outlets including Invezz, Sify, AzerNews, and thecradle.co, turning a story that had been circulating as a Bloomberg-sourced rumor into an on-the-record admission from inside OpenAI itself.
The talks reportedly center on a shared industry standards body for frontier AI models, an idea that has been discussed in working-group meetings since July 2026. What makes the story notable is not that three AI labs are talking about safety, since OpenAI and Anthropic have run joint testing before, but that OpenAI is now willing to say so publicly, in the same week that its own chief scientist and chief executive made separate public comments pointing toward the same effort, and just a day after OpenAI itself launched its next flagship models, GPT-6 Sol and GPT-6 Luna, on September 29, 2026, at API prices roughly 50% below the GPT-5.6 promotions that preceded them.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
OpenAI Confirms Direct Talks With Anthropic and Google DeepMind
According to reporting picked up by Reuters from Bloomberg News, OpenAI’s chief global affairs officer, Chris Lehane, told reporters that the company has been engaging with Anthropic and Google DeepMind on AI safety coordination for several weeks. “It’s better to try to work together to prioritize safety,” Lehane said, according to Reuters. That framing matters because it positions the discussions as risk management between competitors rather than a business partnership, a distinction that becomes important once you consider that all three companies are racing for the same enterprise customers, the same cloud capacity, and the same regulatory goodwill.
Lehane also addressed a question that has hovered over any cross-lab cooperation: whether OpenAI, Anthropic, and Google need special antitrust clearance to coordinate safety practices. He said OpenAI does not believe such a waiver is necessary, arguing that safety-focused information sharing falls outside the kind of coordination that competition law is designed to police. That is a notable legal position for a company that competes head-to-head with both labs on pricing, model releases, and enterprise contracts.
OpenAI’s chief scientist, Jakub Pachocki, backed up Lehane’s comments with his own remarks reported by CNN. “I believe that shared safety standards and international coordination on further AI development need to be priorities now,” Pachocki said, according to CNN. He went further, describing concrete outreach: “We’re talking to some external organizations about potential concrete standards we could put in place.” Sam Altman weighed in too, writing that OpenAI looks “forward to collaborating with our colleagues across the industry to formulate the best version of industry standards,” per the same CNN report.
Timeline: How the Safety Talks Came Together Since July 2026
The public confirmation on September 15 did not appear out of nowhere. Reporting attributed to The Information indicates that working-group meetings among OpenAI, Anthropic, and Google DeepMind staff began quietly in July 2026, months before either side would acknowledge them on the record. Those meetings reportedly focused on the mechanics of a shared standards body rather than on any single incident, suggesting the effort was proactive rather than reactive to a specific failure or breach.
The pace picked up in early September. On September 12, Sam Altman gave an interview with Fortune in which he suggested that leading AI CEOs, naming Dario Amodei and Demis Hassabis specifically, might “get in a room and come up with a plan” to address AI safety collectively. Asked directly whether such a pact could happen, Altman replied simply, “I think that will happen.” Days later, Dario Amodei published an essay arguing that “we must slow the pace at which we improve the capabilities of AI models,” calling for coordination across the US industry and internationally. By September 13 and 14, the standards-body talks were being reported by multiple outlets, and OpenAI’s confirmation followed on September 15. That is a compressed, five-day arc from a CEO hinting at cooperation to on-record confirmation of it, which is fast even by the standards of a news cycle that moves as quickly as AI coverage does.
Demis Hassabis’s FINRA-Style Blueprint for a Frontier AI Standards Body
The specific shape of the proposed body traces back to Google DeepMind chief executive Demis Hassabis, who has floated a US-led Frontier AI Standards Body modeled on the Financial Industry Regulatory Authority, or FINRA. FINRA is not a government agency. It is an industry-funded self-regulatory organization that brokers and exchanges answer to, with the power to write rules, examine members, and refer violations to the Securities and Exchange Commission. Applying that model to AI would mean OpenAI, Anthropic, Google DeepMind, and any other participating lab funding and staffing an independent body that could set testing protocols, conduct evaluations, and flag problems before models ship.
The appeal of a FINRA-style structure is that it lets the industry claim self-governance without waiting on Congress, which has moved slowly on broad AI legislation even as individual states pass their own rules. The risk, which critics of self-regulatory bodies in other industries have raised for decades, is that an organization funded and staffed by the companies it oversees can end up softer than an independent regulator would be. Whether a Frontier AI Standards Body avoids that trap will depend heavily on details that have not yet been made public, including who sits on it, how it is funded, and what happens when a member company disagrees with a finding.
Anthropic’s New Commitment: Independent Evaluators Get Employee-Level Access
Anthropic did not just talk about slowing down. As part of the broader push described in Amodei’s September essay, the company committed to giving independent evaluators permanent, employee-level access inside Anthropic itself, rather than the more limited, time-boxed access that has typically characterized third-party AI audits. That is a meaningfully different commitment than a one-off red-teaming exercise, and it follows Anthropic’s own September 2026 disclosure that it had spent eight months disrupting threat-actor operations that attempted to abuse Claude, a track record the company has pointed to as evidence its safety commitments extend beyond talk. Employee-level access implies evaluators can see internal systems, documentation, and decision-making in something closer to real time, which is the kind of transparency that outside researchers have been requesting from frontier labs for years without much success.
It also raises the bar for OpenAI and Google DeepMind. If Anthropic’s move becomes the reference point for what “serious” safety cooperation looks like, the other two labs will face pressure to match it or explain why they are not. Readers tracking the broader debate over whether AI leaders are willing to slow themselves down can see the wider context in our earlier coverage of OpenAI and Anthropic CEOs’ three-step plan for an AI development brake and the more skeptical response detailed in our report on David Sacks rejecting a proposed AI slowdown deal.
What a Joint AI Safety Standards Body Would Actually Cover
Based on the reporting so far, the proposed standards body would focus on three functions: independent evaluations of frontier models before they reach the public, standardized pre-release safety reviews, and shared risk-assessment frameworks that would apply across participating labs rather than being written separately by each company’s internal safety team. None of that is entirely new in concept. What would be new is having OpenAI, Anthropic, and Google DeepMind agree on a common bar rather than each publishing its own model card with its own testing methodology, which is largely how things work today.
A shared bar has practical value for enterprise buyers too. Procurement teams evaluating OpenAI’s newly launched GPT-6 Sol, priced at $2 per million input tokens and $10 per million output tokens as of September 2026, a steep drop from the $5/$30 pricing OpenAI charged for GPT-5.6 Sol when it launched on July 9, 2026, against Claude or Gemini currently have to reconcile safety disclosures written in different formats with different thresholds, which makes apples-to-apples comparison difficult. A common standards framework, even a voluntary one, would give buyers a more consistent way to compare frontier models on safety grounds rather than relying solely on benchmark scores or list prices. For more on how the current generation of frontier models stacks up, see our breakdown of how to try GPT-6 Astra following its launch.
Not the First Time: The Frontier Model Forum and Prior Cross-Lab Safety Work
Cross-lab safety cooperation did not begin this month. The Frontier Model Forum, founded in 2023 by Anthropic, Google, Microsoft, and OpenAI, already exists as an industry body focused on frontier AI safety research and best-practice sharing. What is different about the September 2026 talks is the specificity: rather than general best-practice sharing, the reported discussions are aimed at a formal standards body with testing authority, closer to a regulator-adjacent structure than a research consortium. The Forum laid the institutional groundwork. The new talks appear to be an attempt to build something with more teeth on top of it.
There is also a direct technical precedent. In summer 2025, OpenAI and Anthropic ran a joint alignment evaluation exercise in which each company applied its own internal safety and misalignment tests to the other’s publicly released models, then published the combined results in an OpenAI blog post dated August 27, 2025. That exercise proved the two companies could share testing methodology and results without either side treating it as a competitive risk, which is likely part of why the September 2026 standards-body talks were able to move as quickly as they did.
CAISI’s Pre-Deployment Review Program Now Covers Five Frontier Labs
Government-linked testing has also expanded quietly over the past year. By September 2025, both OpenAI and Anthropic had agreements with the US Center for AI Standards and Innovation, or CAISI, part of the National Institute of Standards and Technology’s broader AI standards work, giving the government early access to frontier models before public deployment for independent review. On May 5, 2026, Bloomberg reported that Google DeepMind, Microsoft, and xAI joined the same program, bringing the total number of frontier labs participating in CAISI’s pre-deployment review to five. That expansion is a useful data point for anyone trying to gauge whether the September 2026 standards-body talks are a genuine policy shift or just a public-relations exercise: government-run pre-deployment testing across five major labs was already happening before this week’s headlines, which suggests the industry-led talks are building on an existing foundation rather than starting from zero.
For context on how the current administration has approached AI oversight more broadly, see our coverage of the coalition pushing the White House to disclose its AI policy framework, which touches on many of the same tensions between voluntary industry standards and formal government rules.
The Vulnerabilities Government Testers Already Found
CAISI’s evaluations were not just a formality. Testing under the program identified specific, concrete issues in systems from both OpenAI and Anthropic. In OpenAI’s ChatGPT Agent, evaluators found session-scoped remote control and user impersonation vulnerabilities. In Anthropic’s Constitutional Classifiers, the safety layer designed to filter harmful outputs, testers found prompt-injection flaws, cipher-based evasion techniques, universal jailbreaks, and automated attack optimization paths that could bypass the classifiers systematically rather than through one-off tricks.
Those findings matter for two reasons. First, they show that even labs widely regarded as safety-focused, Anthropic in particular given its founding mission, ship systems with exploitable gaps that only surface under independent, adversarial testing rather than internal review. Second, they give the current standards-body talks a concrete justification beyond abstract risk. If a shared testing body existed already, findings like the Constitutional Classifiers bypass techniques could in theory be caught and addressed across all participating labs simultaneously, rather than one company discovering a category of flaw that a competitor’s system shares but has not yet found.
Market and Industry Impact
The immediate market reaction to the confirmation has been muted compared with product announcements, which is typical for safety-policy news, but the longer-term implications for enterprise AI buyers are real. Companies evaluating which frontier model to standardize on for regulated use cases, in finance, healthcare, or government contracting, have been asking for clearer, comparable safety documentation for more than a year. A shared standards body, even a voluntary one without legal enforcement power, gives procurement and compliance teams a reference point that does not currently exist. That could accelerate enterprise adoption in exactly the sectors that have been most cautious about deploying frontier models, since compliance teams often need a defensible external standard to point to before signing off on AI deployment.
There is a competitive angle too. OpenAI, Anthropic, and Google DeepMind are simultaneously racing each other on model capability, measured in benchmark scores, context windows, and pricing, while now also signaling they want to cooperate on the downside risks of that same race. That is not a contradiction so much as a recognition that a serious safety incident at any one lab could trigger regulatory backlash that hits the entire industry. Shared standards function partly as insurance against that scenario, spreading reputational risk across the group rather than leaving any single company exposed when something goes wrong.
Competitive Landscape: Cooperation on Safety, Competition on Capability
It is worth being precise about what is and is not converging here. OpenAI, Anthropic, and Google DeepMind are not merging product roadmaps, pricing, or research priorities. GPT-6 Sol, Claude, and Google’s Gemini 3.8 Flash, which Google priced at $0.75 per million input tokens and $3.75 per million output tokens in September 2026 ahead of a planned increase to $1.50/$7.50 on January 1, 2027, remain separately developed, separately priced, and aggressively marketed against each other. What is converging is narrower: shared testing protocols, shared evaluator access, and a shared body for setting pre-release review standards. Readers comparing the current generation of frontier models on capability rather than policy can find that comparison in our analysis of an OpenAI scientist’s risk warnings following the Astra launch, which lays out how the safety conversation and the product race are running on parallel tracks inside the same company.
The distinction matters for anyone trying to predict how this plays out. Safety cooperation among competitors is not unprecedented, airlines cooperate on safety standards through bodies like the International Air Transport Association while competing fiercely on routes and pricing, and pharmaceutical companies cooperate on clinical trial standards while competing on drug development. The AI industry appears to be reaching for a similar split: cooperate on the floor, compete on the ceiling.
Historical Context: Self-Regulation Playbooks From Other Industries
The FINRA comparison that Hassabis has drawn is instructive precisely because FINRA’s history is mixed. It emerged in 2007 from the merger of NASD and the enforcement arm of the New York Stock Exchange, largely because the financial industry preferred a self-funded regulator to a fully government-run one, and it has both caught significant misconduct and drawn criticism for being too close to the firms it oversees. Research organizations tracking AI governance, including Georgetown’s Center for Security and Emerging Technology, have written extensively about the tradeoffs of industry self-regulation versus statutory oversight for emerging technologies, and those tradeoffs apply directly here.
The other historical reference point is the Asilomar Conference on Recombinant DNA in 1975, where biologists voluntarily paused certain genetic engineering experiments and set safety guidelines before governments had written comparable rules. That episode is frequently cited by AI safety researchers as the model for what industry-led coordination can look like when it works: scientists recognizing risk before regulators catch up, and setting workable rules that later informed formal government policy. Whether the OpenAI-Anthropic-Google DeepMind talks follow that trajectory or drift toward something closer to FINRA’s more contested legacy will depend on decisions that have not been made public yet.
Comparing the Three Labs’ Safety Commitments
| Lab | CAISI Pre-Deployment Access | Standards Body Talks | Notable 2026 Safety Move |
|---|---|---|---|
| OpenAI | Yes, since Sept. 2025 | Confirmed, Sept. 15 | Public confirmation of cross-lab talks via Chris Lehane |
| Anthropic | Yes, since Sept. 2025 | Reported participant | Employee-level access for independent evaluators |
| Google DeepMind | Joined May 2026 | Reported participant, proposal originator | Demis Hassabis’s FINRA-style standards body proposal |
| Microsoft | Joined May 2026 | Not reported as a talks participant | No standards-body role reported as of Sept. 2026 |
| xAI | Joined May 2026 | Not reported as a talks participant | No standards-body role reported as of Sept. 2026 |
Timeline of AI Safety Coordination, 2025 to 2026
| Date | Event |
|---|---|
| Aug. 27, 2025 | OpenAI publishes results of joint alignment evaluation with Anthropic |
| Sept. 2025 | OpenAI and Anthropic establish CAISI pre-deployment review agreements |
| May 5, 2026 | Google DeepMind, Microsoft, and xAI join CAISI, per Bloomberg |
| July 2026 | Working-group meetings begin among OpenAI, Anthropic, and Google DeepMind on standards body |
| Sept. 12, 2026 | Sam Altman signals CEOs may coordinate on an AI safety plan |
| Sept. 13-14, 2026 | Standards-body talks reported publicly, per The Information |
| Sept. 15, 2026 | OpenAI confirms talks via Chris Lehane, Jakub Pachocki, and Sam Altman |
What Comes Next: Five Predictions
First, expect a formal announcement of the standards body’s structure, likely including founding members and initial scope, within the next two to three months, given that working groups have already been meeting since July. Second, Microsoft and xAI are plausible candidates to join given their existing CAISI participation, though neither has been reported as part of the September talks specifically, so their inclusion is not guaranteed.
Third, expect skepticism from policy researchers and some lawmakers who will draw the FINRA comparison critically rather than favorably, questioning whether an industry-funded body can meaningfully constrain the companies that fund it. Fourth, enterprise AI vendors and procurement platforms are likely to start referencing whatever standard emerges as a selling point, similar to how SOC 2 compliance became a baseline expectation in enterprise software sales. Fifth, and most speculatively, the pace of this cooperation will likely track the pace of the underlying capability race. If a new frontier model release triggers a fresh wave of public concern, expect the standards-body timeline to accelerate. If the news cycle moves on to the next product launch instead, expect it to slip, as voluntary industry initiatives often do once the spotlight moves elsewhere.
Why This Story Is Different From Past Safety Pledges
AI labs have made safety pledges before, including the White House-brokered voluntary commitments in 2023 and various signed statements on catastrophic risk. What sets the September 2026 talks apart is the combination of three things happening at once: a named senior executive confirming ongoing talks on the record, a specific structural proposal with a real-world analogue in Hassabis’s FINRA comparison, and prior technical proof of concept in the form of the 2025 joint alignment evaluation and CAISI’s five-lab testing program. Pledges are cheap. A working group that has been meeting since July, building on a testing relationship that already produced published results, is a different category of signal, and it is why this story has moved from rumor to confirmed fact within a single week.
None of that guarantees the standards body will launch, or that it will have any real enforcement power if it does. Industry watchers tracking similar debates around AI valuation and risk pricing can see how quickly sentiment shifts in our coverage of Jim Cramer’s Anthropic IPO enthusiasm running into his own researcher’s 10% AI risk estimate, a reminder that even people inside and around these companies do not agree on how much danger the technology poses.
Frequently Asked Questions
What exactly did OpenAI confirm on September 15, 2026?
OpenAI’s chief global affairs officer, Chris Lehane, confirmed the company has been engaging with Anthropic and Google DeepMind on AI safety coordination for several weeks, including talks about a shared industry standards body, according to Reuters and Bloomberg News reporting.
Is this a formal partnership or merger between OpenAI, Anthropic, and Google?
No. The talks are limited to safety coordination, specifically a proposed shared standards body for testing and evaluating frontier AI models. The three companies remain separate, competing businesses on products, pricing, and research roadmaps.
What is the FINRA-style standards body Demis Hassabis proposed?
Hassabis proposed a US-led Frontier AI Standards Body modeled on the Financial Industry Regulatory Authority, an industry-funded, self-regulatory organization that sets rules and conducts oversight for its member firms rather than a government agency doing so directly.
Does OpenAI think it needs antitrust clearance to coordinate with competitors on safety?
No. Chris Lehane said OpenAI does not believe an antitrust waiver is necessary for the three companies to coordinate on AI safety matters, framing it as risk management rather than anticompetitive coordination.
What is CAISI and how does it relate to this story?
CAISI is the US Center for AI Standards and Innovation. It runs a pre-deployment review program that gives the government early access to frontier AI models before public release. OpenAI and Anthropic joined in September 2025, and Google DeepMind, Microsoft, and xAI joined in May 2026, bringing the total to five participating labs.
Has OpenAI and Anthropic worked together on safety before this?
Yes. In summer 2025, the two companies ran a joint alignment evaluation, applying their internal safety tests to each other’s publicly released models, and published the combined findings in an OpenAI blog post on August 27, 2025.
What vulnerabilities have government testers already found in these companies’ AI systems?
Under CAISI evaluations, testers found session-scoped remote control and user impersonation flaws in OpenAI’s ChatGPT Agent, and prompt-injection flaws, cipher-based evasion, universal jailbreaks, and automated attack optimization paths in Anthropic’s Constitutional Classifiers.
When might a formal AI safety standards body actually launch?
No launch date has been announced. Working groups have reportedly been meeting since July 2026, and given that pace, a formal structure with named founding members could plausibly be announced within the next few months, though nothing has been confirmed publicly.
