OpenAI, Google and Anthropic have been holding discussions about creating an industry body that could establish common safety standards for advanced artificial intelligence, marking an unusual area of cooperation among companies competing intensely to build the world’s most powerful AI systems.
Representatives from the three companies have reportedly been meeting for months to discuss a framework covering areas including model testing and independent safety audits. The talks remain preliminary, and no formal organization or binding agreement has been announced.
The discussions have taken on greater significance as AI executives increasingly warn that advances in frontier models may be moving faster than existing safeguards.
Anthropic CEO Dario Amodei has called for leading AI developers to coordinate on slowing the pace of development when necessary to give safety systems time to catch up. OpenAI CEO Sam Altman and Google DeepMind’s Demis Hassabis have expressed support for greater coordination around frontier AI safety.
The companies have previously experimented with cooperation. OpenAI and Anthropic conducted cross-company safety evaluations in 2025, allowing each laboratory to test the other’s models for potentially dangerous or misaligned behavior.
The stakes are rising as frontier models gain stronger abilities in cybersecurity, autonomous research and other sensitive areas. OpenAI recently classified its GPT-6 Astra model at its highest cybersecurity capability level and has called for national requirements governing increasingly capable AI systems.
A joint standards body would represent a significant shift for an industry built around rapid competitive development. But major questions remain over who would enforce the standards, whether participation would be voluntary and whether governments would accept an industry-led system as sufficient oversight.
For now, the talks signal something unusual: some of the companies racing hardest to advance artificial intelligence are increasingly discussing whether that race itself needs rules.
