Google, OpenAI, Anthropic Line Up SAFA AI Watchdog

The three frontier labs are shaping a self-run Standards Authority for Frontier AI, even as California readies its own AI regulator.

According to insider reports from The Information cited by heise online, plans for a supervisory authority for AI systems organized by Google, Anthropic and OpenAI are increasingly taking concrete shape, under the circulating name Standards Authority for Frontier AI (SAFA). The heise report says the body would define rules for how developers report safety incidents, support independent third-party auditors of AI models, set auditor qualifications, and specify how the companies' self-commitments work in practice — while it remains unclear whether SAFA would audit AI systems itself.

The unusual part is who is driving it: The Associated Press, in The Hindu's report, notes the CEOs of Anthropic and OpenAI have declared their most powerful models dangerous and in need of independent testing — a rare instance of unity between rivals, delivered in essays and speeches to the United Nations. Forkast reports that on September 23 Anthropic, Google and OpenAI launched SAFA as a FINRA-style self-regulatory body that Meta, xAI and Nvidia publicly opposed, meaning the labs are splitting the industry over how frontier AI should be governed.

Skeptics say the labs' alarm is self-serving. The Hindu, citing AP reporting, quotes former OpenAI geopolitics lead Sarah Shoker as warning that focusing on existential risk deprioritizes present harms like AI in warfare, while PitchBook analyst Harrison Rolfes argues safety messaging builds a 'wall or moat' ahead of IPOs and midterms. Conrad Stosz, a former CAISI leader now at Transluce, tells The Hindu it is 'a little ambiguous what embedded evaluators means,' questioning whether auditors can investigate without losing independence.

heise reports names in circulation for SAFA leadership include Sriram Krishnan, a former venture capitalist and Trump administration AI policy advisor, and Arati Prabhakar, former Biden-era White House science chief, for CEO, with Condoleezza Rice and David Friedberg discussed for chairman, and METR's Beth Barnes and Paul Christiano as scientific advisors. The Hindu notes President Trump has dismissed AI risk talk as a 'HOAX,' while California's governor explicitly backs AI rules, and Forkast observes an antitrust suit filed September 18 alleges safety coordination among competitors is a cartel.

heise points out that California already plans its own supervisory authority for AI companies, so the labs' claim they act only because state bodies reject regulation is, in its words, not entirely accurate. Forkast reports Anthropic faces a self-imposed September 30 deadline for Phase 1 of its provable-inference prototype, and that the industry is struggling with containment after OpenAI's second training halt on September 20. Whether SAFA publishes standards, staffs up, or stays silent in the IPO rush will show if self-regulation is credible or a moat.