After reading this, the reader knows OpenAI, Anthropic and Google are reportedly discussing AI safety, while the scope and outcome remain unconfirmed.

AI labs reportedly open joint safety talks

A report says three leading model developers have held discussions for weeks. None had publicly confirmed a concrete agreement when this article was prepared.

OpenAI is reportedly in talks with Anthropic and Google about cooperation on AI safety, according to a Bloomberg report relayed by Reuters on September 15.

In plain terms, commercial rivals may be looking for areas where a common safety baseline is more useful than separate company rules. The important word is “may”: Reuters said it could not independently verify the talks, and the companies did not immediately confirm them in the report.

Why it matters

The largest model developers face many of the same hard-to-contain risks. A severe vulnerability, dangerous capability or weak evaluation practice at one company can affect users, suppliers and regulators beyond that company’s products. Shared testing methods or disclosure channels could reduce gaps between labs.

Cooperation also has limits. The companies compete for customers, researchers and computing capacity. Any joint effort must avoid exchanging commercially sensitive information or turning voluntary standards into a barrier for smaller developers. A discussion is therefore far from a shared technical regime.

Reuters, citing Bloomberg, reported that the conversations had continued for several weeks. OpenAI chief global affairs officer Chris Lehane was quoted as saying safety is an area where companies should cooperate and that the discussions would not require an antitrust waiver. Those statements are attributed reporting rather than a public transcript or joint announcement.

The report did not identify a working group, timetable, written charter or agreed technical deliverable. It also did not say whether Meta, xAI, independent researchers, governments or civil-society groups would participate. Without those details, the most accurate description is preliminary reported contact among three companies.

Evidence will be the test

Useful cooperation could take several forms: common definitions for severe incidents, compatible evaluation protocols, protected channels for disclosing vulnerabilities, or joint preparation for threats that cross product boundaries. Each option creates different questions about transparency and enforcement.

OpenAI has separately argued for stronger safety evidence, shared standards and durable public policy as AI capabilities rise. That public position is consistent with the reported talks, but it does not confirm that the conversations occurred or that Anthropic and Google accepted any proposal.

Company-led standards can move faster than legislation and draw on direct technical access. They can also let the participants choose the tests by which they are judged. Credible work would publish methods, explain failure thresholds and invite independent scrutiny rather than relying on general commitments.

The next meaningful development is a primary document: a joint statement, named participants, a specific evaluation or incident-sharing protocol, and dates for publication or review. Until then, the story is about an apparent willingness to talk, not a safety pact.

Verification

  1. UNVERIFIED — OpenAI, Anthropic and Google have held AI-safety cooperation talks for several weeks. Reuters relayed Bloomberg’s report and said it could not independently verify it; no joint primary statement was located. Via: https://www.reuters.com/technology/openai-is-working-with-anthropic-google-ai-safety-bloomberg-news-reports-2026-09-15/
  2. UNVERIFIED — Chris Lehane said the companies should cooperate on safety and that the talks require no antitrust waiver. The remarks were attributed by Bloomberg and relayed by Reuters; no public transcript was located. Via: https://www.reuters.com/technology/openai-is-working-with-anthropic-google-ai-safety-bloomberg-news-reports-2026-09-15/
  3. VERIFIED — OpenAI has publicly advocated stronger safety evidence, shared standards and policy action as AI capabilities increase. Primary source: https://openai.com/
  4. VERIFIED — Reuters reported that OpenAI, Anthropic and Alphabet did not immediately comment and that Reuters could not verify the Bloomberg report. Source: https://www.reuters.com/technology/openai-is-working-with-anthropic-google-ai-safety-bloomberg-news-reports-2026-09-15/
  5. PARTIALLY VERIFIED — Common evaluations or incident channels could reduce cross-company safety gaps. This is an analytical inference; no specific joint mechanism has been announced.
  6. UNVERIFIED — The talks will produce a joint standard, agreement or operational program. No such outcome had been announced when this article was prepared.

Glossary candidates

  • Safety evaluation
  • Incident disclosure
  • Antitrust waiver
  • Voluntary standard
  • Capability threshold

Cold-reader sentence: OpenAI, Anthropic and Google are reportedly discussing AI safety, but no company had confirmed a joint program or deliverable.