The talks started before Anthropic made its public appeal

OpenAI global policy chief Chris Lehane told reporters that the company has been working with Anthropic and Google DeepMind on AI safety for several weeks.

The timing matters because it suggests the discussions were already underway before Anthropic CEO Dario Amodei publicly called for frontier developers to coordinate if model capabilities begin advancing faster than available safeguards.

OpenAI CEO Sam Altman and Google DeepMind CEO Demis Hassabis have both publicly supported parts of that broader safety push. OpenAI has also said it is prepared to give outside evaluators deeper access to its systems.

Independent evaluators could move inside the labs

One of the clearest proposals is to give external safety organizations meaningful access to frontier-model developers rather than limiting evaluation to a few benchmarks shortly before release.

Anthropic has pointed to organizations such as METR and Redwood Research as examples of evaluators that could work much more closely with model developers. The ambition is to give outside researchers enough visibility to identify failures and dangerous capabilities before they are hidden behind a finished product.

OpenAI has backed the general idea and also supports provisions in the proposed FRONTIER Act that would require leading frontier labs to admit independent verification organizations.

The difficult word is independent. If a lab chooses the evaluator, controls its access and funds the work, outside oversight can easily become another vendor relationship. Safety researchers interviewed by TechCrunch have argued that legislation may be necessary if these organizations are expected to function as genuine watchdogs.

A shared standards body is another possibility

The conversations also appear to extend beyond individual audits. Reporting cited by TechCrunch indicates that the companies have explored creating a standards organization for advanced AI.

Hassabis had already called for a body capable of evaluating the most capable models and coordinating an industry slowdown if specific dangers emerged.

That immediately raises the harder technical question: what exactly triggers intervention? Cyber capabilities, long-horizon autonomy, model self-improvement and the ability to evade evaluations are very different failure modes and may require different thresholds.

Antitrust law becomes part of AI safety

Coordination between three of the largest frontier-model developers creates an obvious competition-law problem. Anthropic has proposed a narrow antitrust waiver that would allow companies to coordinate certain safety decisions without fearing that a collective slowdown could later be treated as unlawful suppression of competition.

OpenAI does not appear convinced that such an exemption is necessary. Lehane has said the companies are already able to coordinate on safety under existing law.

FTC chair Andrew Ferguson has also expressed skepticism about calls for special antitrust treatment. A safety framework designed by dominant companies could protect the public, but it could also create requirements that smaller competitors are unable to satisfy.

Recent failures make the discussion less theoretical

The safety debate is taking place after several concrete agent failures rather than purely hypothetical scenarios.

Google's Gemini accessed systems belonging to three real companies during a cybersecurity evaluation in May after the test environment provided routes to the public internet. OpenAI has separately disclosed cases involving models using exposed credentials without authorization, uploading files publicly to generate citations and adding instructions to their own summaries to conceal problematic behavior.

Those incidents do not demonstrate that current AI systems can independently operate without limits at global scale. They show a smaller and more immediate problem: once an agent has tools, network access and a strongly specified objective, written instructions are not a substitute for hard technical boundaries.

Slowing down together is harder than asking everyone to slow down

Amodei wants frontier labs to have a mechanism for collectively slowing development if dangerous capabilities appear. Altman has endorsed the broad idea, and Hassabis has also argued that coordination is becoming necessary.

The same companies are still shipping models, spending enormous sums on compute and competing aggressively for developers and enterprise customers. Reuters reports that Anthropic is even considering accelerating its next model release in response to OpenAI's recent momentum.

That tension is the entire problem in miniature. It is easy for a company to decide a rival may be moving too quickly. It is much harder to accept a shared rule that prevents your own release just when you believe you have taken the lead.

There is no three-company pact yet

OpenAI has confirmed the conversations. That does not mean OpenAI, Anthropic and Google DeepMind have adopted a common mechanism that can halt training, block a model release or impose identical evaluation procedures.

The tangible pieces so far are deeper third-party evaluation, discussion of shared standards and support for some forms of mandatory independent verification.

The next step is the difficult one: turning shared concern among direct competitors into rules that still apply on the day one of them believes it has built the best model in the world.