OpenAI, Google and Anthropic — the three companies leading the AI race — have been meeting regularly since July to discuss building a joint body that would test and audit frontier models before release. The talks were first reported by The Information and then confirmed by the Washington Post and several other outlets.
The idea traces back to Google DeepMind chief Demis Hassabis, who published an essay on July 14 proposing a FINRA-style organization for frontier AI — a reference to the body that oversees US securities brokers. Its job would be to independently check models for risks before they reach the public: cybersecurity threats, biological misuse, manipulative behavior.
Working groups from the three labs, made up of executives below CEO level, have been hashing out the details since summer: how often audits should happen, who should run them, and what counts as safe enough to ship. OpenAI's Sam Altman has said he backs a testing and auditing organization, but argues the major labs will need to build it themselves rather than wait for Washington.
The report landed just after a weekend in which Anthropic's Dario Amodei published an essay calling for a slower pace of AI development — and Altman, Elon Musk and Hassabis all publicly agreed with him. Tech stocks dipped on the news, as investors read the united front from rival CEOs as a sign the risks are more serious than markets had priced in.
Skeptics see real problems ahead. An industry-run safety body risks repeating the credit-rating agencies' pre-2008 conflict of interest, where the companies being graded also paid the graders. By 2028 there could be 14 to 16 frontier models competing at once, which means a constant stream of reviews, and compliance costs would likely hit smaller labs harder than the giants funding the effort. There's also the question of how a US-based body would mesh with the EU's existing General-Purpose AI Code of Practice.
Earlier self-regulation attempts, like the 2023 Frontier Model Forum, never amounted to much. Whether companies that compete daily for customers and talent can actually agree on shared rules this time remains an open question.
