Three of the world's leading AI developers — Google, OpenAI and Anthropic — are planning an industry-run safety organisation called the Standards Authority for Frontier AI (SAFA), according to multiple reports this week. The body would set testing and auditing standards for the most powerful AI models, with a target launch in late 2026 or early 2027 — but without direct government oversight, a design that is already drawing criticism.
Details emerged through reporting by The Information on September 24 and follow-up coverage from PYMNTS, TheStreet and others. The companies have not issued a formal joint announcement, and many specifics, including leadership and governance structure, remain unsettled.
What SAFA Would Do
As described in reports, SAFA would function as a formal standards body for frontier AI rather than a research forum. Its expected remit includes:
- Third-party safety testing of frontier models before public release
- Incident-reporting rules covering serious failures and misuse
- Qualification standards for auditors who evaluate AI systems
That would make SAFA more muscular than the Frontier Model Forum, the group Anthropic, Google, Microsoft and OpenAI founded in 2023 to share safety research and best practices. SAFA is pitched as something closer to a regulator — one funded and run by the industry it oversees.
Several outlets report the companies have approached Sriram Krishnan, a former Andreessen Horowitz partner and former senior White House AI policy adviser, about serving as chief executive.
A FINRA Model for AI
The concept traces back to a proposal from Google DeepMind CEO Demis Hassabis for a US-led frontier AI standards body modelled on FINRA, the industry-funded organisation that writes rules for and supervises broker-dealers under the oversight of US securities regulators.
According to reports, the labs initially favoured a public-private partnership with federal oversight, but shifted toward self-regulation after a draft White House executive order stalled. In mid-September, OpenAI's chief global affairs officer Chris Lehane confirmed the company had spent weeks talking with Anthropic and Google DeepMind about safety coordination, including a joint standards body. He argued the companies did not need an antitrust waiver, describing the effort as risk management rather than anticompetitive collusion.
The Independence Problem
Critics were quick to point out the obvious tension: a safety authority created, funded and staffed by the firms it is meant to police. Commentators warned the arrangement could entrench the largest labs, set standards that smaller rivals struggle to meet, and lack the independence needed to block a risky release.
Supporters counter that no comprehensive federal AI law exists, and that binding shared testing standards — even industry-run ones — are better than today's patchwork of voluntary commitments. Financial markets offer a precedent: FINRA's self-regulatory model has operated for decades, though always under a government regulator's supervision.
Why It Matters
SAFA arrives at a moment of intense pressure on AI governance. In recent weeks, reports of agents escaping sandboxes, unauthorised system access and destructive coding-agent errors have sharpened concerns about frontier-model safety. On September 23, OpenAI's Sam Altman and Anthropic's Dario Amodei addressed the UN Security Council, urging world leaders to adopt shared AI safety standards before the technology outpaces institutions.
At the same time, lawmakers are moving on their own. Members of Congress have proposed restricting advanced AI development until a dedicated federal agency exists, California's governor has ordered work on independent audits and an AI kill switch, and cities such as New York are debating their own rules. An industry-built standards authority could shape — or pre-empt — those efforts.
For businesses deploying AI, SAFA could eventually become a de facto benchmark. If its testing and incident-reporting standards gain traction, enterprise buyers and insurers may start asking whether a model has passed SAFA-style audits, much as they now ask about SOC 2 or ISO certifications.
What to Watch
The critical questions are governance and teeth. Watch whether SAFA includes independent board members, publishes its test results, and has any power to delay a model release. Also watch whether other major developers — Meta, xAI, Microsoft and leading Chinese labs — join, and whether Washington chooses to formalise oversight of the body, as the SEC does with FINRA.
