OpenAI confirms safety talks with Anthropic and Google

OpenAI has been talking safety with rivals Anthropic and Google DeepMind for weeks. That is no longer a second-hand report from The Information. OpenAI has said it on the record.

A spokesperson confirmed it to CNBC on Tuesday. Global policy chief Chris Lehane told reporters in Washington, first reported by Bloomberg and carried by TechCrunch. The talks have been running since Demis Hassabis in July called for a U.S. standards body modelled on FINRA, CNBC reported.

The news is not the weekend essay. It is that the three largest U.S. labs now confirm they are coordinating, while they disagree on antitrust, statute and pace.

For CIOs, CISOs and boards this is vendor governance. Evaluators may get employee-like access. Standards can harden into procurement defaults. And Nvidia, which sells the capacity, said in public that the industry does not need new laws.

What is new

On Saturday Dario Amodei asked the industry to pace frontier development. Sam Altman, Elon Musk and Hassabis agreed. On Monday China and Donald Trump said no. That is already covered.

Tuesday added three facts.

First: OpenAI confirmed the talks. CNBC got it from a spokesperson. Lehane said the company had been working with Anthropic and Google for several weeks. Google and Anthropic did not immediately respond to CNBC.

Second: OpenAI does not want an antitrust waiver. Amodei had floated a narrow exemption so labs could coordinate on safety without hitting U.S. competition law. Lehane said they do not need one, according to TechCrunch and Reuters’ account of Bloomberg. “It's better to try to work together to prioritize safety,” he said. He compared it to airlines helping each other on safety without that being treated as a cartel.

Altman has already flagged the risk: if the coordination looks like suppressing competition, it can collide with antitrust law. The Information, cited by TechCrunch, reported that the three labs have been working toward a standards body, and that Altman told staff it would need to happen without U.S. government support.

Third: OpenAI is backing a specific legal mandate. Politico reported that Lehane on Tuesday supported a provision in the FRONTIER Act from Representatives Jay Obernolte and Lori Trahan. It would force the largest labs to admit “independent verification organizations,” or IVOs, to check that models are developed safely. The duty would apply only to companies above high revenue and compute thresholds. Politico called it the first time OpenAI has backed a specific federal mandate for third-party assessors.

Lehane sat Monday with one of the sponsors. “We can support that,” he said, according to Politico. The same week California Governor Gavin Newsom signed SB 813, which sets an accreditation path for such evaluators. OpenAI endorsed it late.

Reuters noted it could not independently verify Bloomberg’s briefing. The CNBC on-record confirmation is therefore the strongest public trail. These are still talks, not a signed pact.

Huang and Zuckerberg reject a slowdown

The same day, at Salesforce Dreamforce in San Francisco, the agreement split on stage.

Amodei restated the pacing plan in front of about 12,000 people. Jensen Huang followed. Speed and safety are “a false choice,” he told Marc Benioff, according to CNBC. “You could definitely have both at the same time.” And: “Run as fast as you can.” If the product is not safe, pause. But “we don’t need new laws — we don’t need new regulations.” Safety is an engineering problem, not a legislative one.

Altman, later the same day, rejected conditional responsibility. Companies that say they will only be responsible if others are do not answer what the public is asking, he said. “There should be no qualifier on that.”

Mark Zuckerberg posted a counterargument on Tuesday, Forbes reported: labs have a “natural incentive” to keep models safe because the liability is large. He backed independent evaluators, but framed recursive self-improvement as the wrong race. Meta, he wrote, commits most compute to serving people rather than racing toward self-improving systems.

The labs want a shared pace. The chip and platform owners do not want a shared brake. That is the real split, not the essay.

What it means for European and Norwegian leaders

You buy models from companies that are now negotiating standards with each other. That can mean shared tests, shared incident reporting and shared safety-case requirements before large training runs. On 9 September OpenAI wrote that industry standards should complement, not replace, mandatory federal safeguards, and that it would build a voluntary effort “with or without government support.”

It can also mean lock-in. If OpenAI, Anthropic and Google agree what “safe enough” means, that definition becomes the template copied by buyers, auditors and supervisors. The EU already has the AI Act for systemic risk. The United States does not have a single line. Trump has called the safety fears a “hoax.” China called Amodei’s plan a Cold War move. A standards body without Beijing and without Washington then becomes the de facto rulebook for Western customers.

The IVO track is the most concrete. Amodei and Altman have promised evaluators with employee-like access. The FRONTIER Act would make that a duty for the largest labs. For a Norwegian or EU organisation that means three questions: who reads your data in the evaluation, what the lab can refuse, and what happens to your tenant if an evaluator stops a model.

The antitrust track matters equally. If the talks look like a pace cartel, U.S. and EU enforcers can open a case. If they look like aviation safety, they can become the template. Do not guess. Require in writing what the vendor shares with rivals: test data, vulnerabilities, capacity plans, pricing, or only safety methods.

The Huang track is capacity. If Nvidia and Meta do not slow down while labs promise to pace, you get different clocks for model, cloud and silicon. Contracts that assume a smooth rise in capability can fail both ways: the next model arrives later, or it arrives without the evaluator paperwork the board now wants.

What the board should do now

Require OpenAI, Anthropic and Google to answer in writing whether they take part in safety coordination, what is shared, and whether they believe a FRONTIER Act IVO duty would hit their product line. Ask for the evaluator’s name, access level and what is logged from your tenant.

Write procurement that survives both outcomes. If the United States gets an IVO mandate, you should have a right to see a summary of the evaluation. If it does not, you still need your own red team, an incident clause and a right to switch models without penalty.

Separate pace from control. The EU AI Act, NIS2 and GDPR will not wait for Capitol Hill. Agents that touch code, identity and money need authority, logs and rollback whether Huang runs or Amodei brakes.

Do not read “safety talks” as a pause. Read them as the three labs trying to write the rules before Congress does. The board should know which rules you then inherit.

Sources and media

Primary source: CNBC, “OpenAI, Google, Anthropic discussing collaboration on AI safety issues”, 15 September 2026, https://www.cnbc.com/2026/09/15/open-ai-google-anthropic-safety.html

TechCrunch, “OpenAI, Anthropic, Google have been in talks on AI safety for weeks”, 15 September 2026, https://techcrunch.com/2026/09/15/openai-anthropic-google-have-been-in-talks-on-ai-safety-for-weeks/

Reuters, “OpenAI is working with Anthropic, Google on AI safety, Bloomberg News reports”, 15 September 2026, https://www.reuters.com/technology/openai-is-working-with-anthropic-google-ai-safety-bloomberg-news-reports-2026-09-15/

Politico, “OpenAI backs bipartisan House plan for third-party safety assessments”, 15 September 2026, https://www.politico.com/news/2026/09/15/openai-backs-bipartisan-house-plan-for-third-party-safety-assessments-01076588

CNBC, “Nvidia’s Huang diverges with CEOs of Anthropic, OpenAI on AI safety at Dreamforce”, 15 September 2026, https://www.cnbc.com/2026/09/15/nvidia-and-anthropic-ceos-diverge-on-ai-safety-at-dreamforce.html

OpenAI, Chris Lehane, “The AI policy window is open. We need to act.”, 9 September 2026, https://openai.com/index/ai-policy-window/

Dario Amodei, “We Must Pace the Frontier”, 12 September 2026, https://darioamodei.com/post/we-must-pace-the-frontier

Thumbnail: OpenAI Image 2 / hogby.ai