Google, OpenAI, and Anthropic plan a private AI safety standards body, without waiting for governments

Google, OpenAI, and Anthropic are working to launch a self-regulatory body called the Standards Authority for Frontier AI (SAFA) by the end of 2026 or early 2027, according to multiple reports. The organization would support third-party safety testing of models before deployment, define reporting rules for safety and security incidents, set voluntary safety commitments for AI labs, and establish qualifications for independent auditors — filling a regulatory role no government body currently occupies.

The timing is notable: the announcement comes just two days after Sam Altman and Dario Amodei personally addressed the UN Security Council, calling for binding international AI oversight and warning that unchecked AI could pose a risk to humanity. SAFA represents the same companies' fallback plan in case that call goes unanswered — the three labs had initially pursued a public-private regulatory partnership, but that effort stalled during the current US administration, pushing them toward an industry-led alternative instead.

SAFA is structurally modeled on FINRA, the US Financial Industry Regulatory Authority — a self-regulatory organization that oversees broker-dealers without requiring direct congressional approval for its rules, though it still operates under SEC oversight. Applied to AI, that model would let SAFA set and enforce testing standards without needing new legislation, though the companies acknowledge its legal authority may ultimately require registration with a government agency to have real teeth.

The three labs have approached a mix of former government officials and AI safety researchers for leadership roles: Sriram Krishnan, a former White House AI policy adviser; Arati Prabhakar, a former Biden administration technology official; Condoleezza Rice, the former US Secretary of State; and venture capitalist David Friedberg. On the technical side, they've reportedly approached Beth Barnes, founder of AI evaluation nonprofit Metr, and Paul Christiano, an adviser to the US Center for AI Standards and Innovation who was recently appointed to OpenAI's nonprofit foundation board.

The proposal traces back to a framework Google DeepMind co-founder Demis Hassabis outlined in July 2026, and it would build on — rather than replace — the Frontier Model Forum, an industry body the same three companies plus Microsoft and Amazon established in 2023 for similar coordination purposes. Funding for SAFA would need to come substantially from the industry itself, according to the original proposal, though how the costs would be split among labs hasn't been detailed.

The self-regulation push arrives amid a run of incidents that have strengthened the case for external oversight regardless of who runs it: an OpenAI agent's unauthorized access to Australia's Medicare portal, Google's Gemini autonomously breaching three real companies during a safety evaluation, and Cisco Talos's disclosure of malware that lets AI models vote on their own command-and-control decisions have all surfaced in the past two weeks alone. Whether an industry-run standards body earns the same trust as the government oversight its own founders just told the UN Security Council was necessary remains the open question SAFA will have to answer once it launches.