OpenAI confirmed on September 15, 2026, that it had been coordinating with Anthropic and Google DeepMind on AI safety for several weeks. Chris Lehane, OpenAI’s global policy chief, said the work was intended to prioritize safety and that OpenAI did not believe the coordination required an antitrust waiver. The disclosure is significant—but it describes talks, not a signed agreement or a finished oversight system.
OpenAI confirms weeks of AI-safety coordination
OpenAI is the company that confirmed the discussions. Anthropic and Google DeepMind are described as participants in that coordination, but neither company independently confirmed the talks in the public statements surrounding the disclosure.
The immediate question is whether the three leading AI labs are already bound by a formal safety agreement. The answer is no: the current framework consists of reported coordination, policy proposals and commitments at different stages of development.
That distinction matters because safety cooperation among direct competitors can serve two very different purposes. It can create shared checks for increasingly capable models—or allow the largest companies to shape the rules in ways that make life harder for smaller rivals. The debate is already moving along both tracks.
What the companies are proposing
Anthropic CEO Dario Amodei published his pacing proposal on September 12, 2026. “Pacing” does not mean stopping AI development. It means slowing the rate at which model capabilities improve while continuing technical work.
Amodei proposed embedding independent third-party evaluators inside AI companies. These evaluators would receive ongoing, employee-like access to tools, workspaces and permissions comparable to internal risk teams. Anthropic says it is making that commitment unilaterally. Sam Altman endorsed independent evaluators and said OpenAI would follow suit, but the scope and timing of OpenAI’s implementation have not been announced.
The idea is straightforward in principle: outside teams should be able to examine safety practices, incidents, training pipelines and alignment work from inside the organizations building frontier models. In practice, the hard questions are about access, independence and what happens when an evaluator finds a serious problem.
Google DeepMind CEO Demis Hassabis proposed a different structure on July 14, 2026: a U.S. public-private or self-regulatory organization modeled on the Financial Industry Regulatory Authority, or FINRA. The proposed body could define thresholds for frontier models and develop testing in areas including cybersecurity, biological threats, deception, guardrail bypass and alignment.
Hassabis’s proposal included voluntary sharing of models for testing up to 30 days before release, with possible formalization later. It would require substantial funding, likely from industry, to support technical staff and the computing resources needed for large-scale evaluations. It remains a proposal, not an operating regulator.
Why safety proposals raise competition concerns
The central dispute is not whether AI systems need safeguards. It is who writes the rules, who can participate in the process and who controls the testing infrastructure.
Cohere CEO Aidan Gomez has argued that dominant AI labs should not control the standards for a technology with broad economic and social consequences. Andrew Ferguson, chair of the U.S. Federal Trade Commission, expressed deep skepticism about possible antitrust exemptions and warned that they could create barriers to entry.
OpenAI’s position is narrower: Lehane said the company does not believe an antitrust waiver is necessary for safety coordination. Amodei’s proposal, by contrast, discusses the possibility that some forms of cooperation could require a narrow government waiver. No waiver has been granted.
This is why the phrase “AI-safety coordination” should not be treated as a synonym for either public-interest oversight or anticompetitive behavior. The same arrangement could improve testing while also concentrating influence, depending on who controls the evaluators, the standards and the release thresholds.
Four oversight models on the table
| Proposal | Who evaluates or sets the rules | Scope | Current stage |
| Embedded evaluators | Independent third-party teams inside individual AI labs | Safety practices, incidents, training pipelines and alignment work | Anthropic says it is committing; OpenAI has said it would follow, with scope and timing still unspecified |
| FINRA-like standards body | A proposed public-private or self-regulatory organization with independent technical experts | Pre-release testing and evolving benchmarks for frontier models | Proposed by Demis Hassabis; not formally established |
| Rival testing | Competitor AI companies testing one another’s systems | Dangerous capabilities before release | Proposed by Elon Musk; no named rival has agreed to adopt it |
| Federal independent verification | Independent verification organizations under a proposed federal framework | Safety assessments of leading AI labs | Supported by OpenAI through the proposed FRONTIER Act; the legislation has not been enacted |
The models differ most sharply in who holds the keys. Embedded evaluators would work within each lab. A FINRA-like body would centralize standards and testing. Rival testing would put competing companies in the reviewer’s seat. Federal verification would place the process within a government-backed framework.
Each approach also creates a different independence problem. An evaluator inside a lab may gain valuable access but face institutional pressure. A shared standards body may bring consistency but give incumbent companies outsized influence. Rival testing could expose problems from a competitor’s perspective, while also creating conflicts over proprietary systems and sensitive capabilities.
What happens next
The practical status is therefore mixed. OpenAI has disclosed several weeks of coordination with Anthropic and Google DeepMind. Anthropic has proposed pacing and embedded evaluators. Google DeepMind has proposed a FINRA-like standards body with pre-release testing. OpenAI supports independent verification through a proposed federal measure, while Elon Musk’s rival-testing idea has not been adopted.
The next meaningful test is not another statement of principle. It is whether these proposals produce access rules, independent evaluators, shared benchmarks and clear authority to act when a model fails them. Until then, the AI-safety debate remains a tug-of-war between faster capability development, stronger oversight and the risk that the companies building the frontier will also end up writing its rulebook.
