In the wake of OpenAI’s second training halt and explicit liability signals from the Treasury Secretary, the industry’s largest players have moved to formalize a self-regulatory body: the Standards Authority for Frontier AI. Google, OpenAI, and Anthropic are proposing a structure modeled after the Financial Industry Regulatory Authority – common safety benchmarks, standardized model testing, voluntary safety commitments, auditor qualifications, and incident reporting protocols. The target launch is late 2026 or early 2027. The concept, first attributed to Google DeepMind CEO Demis Hassabis in July 2026, has been under active discussion for weeks.
The timing is not coincidental. SAFA arrives in the same window as OpenAI’s second training halt, Treasury Secretary Bessent’s explicit rejection of liability shields, and an antitrust lawsuit that alleges safety coordination is cartel behavior. The labs are building the gate while the legal system is testing whether the gate itself is the problem.
What SAFA Would Do
The proposed body would establish common testing and risk-assessment practices for frontier models, support third-party testing organizations, set standards for independent auditors, and potentially conduct some model evaluations itself. Several figures have been approached for leadership roles, including former White House AI policy adviser Sriram Krishnan, former Biden administration technology official Arati Prabhakar, former Secretary of State Condoleezza Rice, and venture capitalist David Friedberg.
The scope is broad. But the enforcement question remains open. If SAFA operates on voluntary commitments – which all reporting to date suggests – it creates a structure where compliance is a signal, not a constraint. Labs that choose not to participate face no formal consequence. Labs that do participate gain the credibility of third-party oversight without the accountability that comes with mandatory compliance.
Lehane’s Lobbying and the Policy Window
OpenAI’s Chief Global Affairs Officer, Chris Lehane, confirmed the discussions on September 15 to Bloomberg, CNBC, and TechCrunch. Lehane has been in Washington meeting with lawmakers on catastrophic AI risks. On September 9, he published an OpenAI blog post titled “The AI policy window is open. We need to act.” – a three-pillar framework calling for mandatory national safety requirements, state-level momentum, and industry-led standards.
Lehane also publicly supports a provision in the FRONTIER Act that would require frontier labs to allow “independent verification organizations” into their companies. This positioning is careful: OpenAI is calling for third-party oversight while simultaneously defining the terms of that oversight. The company published its own third-party assessment framework on September 22 – unilaterally setting the methodology for the scrutiny it faces.
Who Is Not at the Table
The industry is far from unified. At Dreamforce on September 15, Meta, xAI, and Nvidia publicly opposed the formation of SAFA. Mark Zuckerberg argued that AI labs already face strong liability incentives and that existing laws are sufficient. Jensen Huang of Nvidia went further, telling Salesforce CEO Marc Benioff that AI laws are “just completely unnecessary” and that model makers should simply take responsibility for their products. xAI opposed alongside both companies.
The opposition is not marginal. Meta and xAI are building frontier models. Nvidia supplies the compute infrastructure. Their refusal to join SAFA means the body would launch without the participation of at least two frontier-capable model builders and the dominant hardware supplier. A safety standards body that excludes its most aggressive competitors is not a safety body – it is a competitive positioning play with safety branding.
The Antitrust Shadow
SAFA arrives in a hostile legal and political environment. On September 15, Senators Hawley and Cruz blocked an antitrust exemption in the NDAA that would have provided legal cover for coordinated safety standards. Three days later, four plaintiffs filed Buist et al. v. Anthropic PBC et al., alleging that safety coordination among competitors constitutes an output-restricting cartel under Section 1 of the Sherman Act. Cohere CEO Aidan Gomez has labeled the proposed body “a cartel by any other name.”
On September 19, President Trump dismissed AI safety as a “hoax” and announced the creation of an AI Force with no safety mandate. The White House has put a draft executive order on hold, telling the companies to reach industry consensus first. The political environment does not merely fail to support self-regulation – it actively questions whether self-regulation is distinguishable from market manipulation.
The Structural Question
The core tension of SAFA is governance. When the largest players in an industry define the safety benchmarks, the risk of regulatory capture is inherent. If SAFA becomes the de facto standard-setter, it creates a compliance barrier for smaller startups that cannot afford the costs of meeting those benchmarks – benchmarks defined by the incumbents who can.
The FINRA comparison is instructive. FINRA succeeded because Congress granted it statutory authority through the Securities Exchange Act. Without equivalent legislative backing, SAFA operates as a voluntary pact among competitors – which is precisely what the antitrust lawsuit alleges is illegal. The body can launch. But without legislative authority, every standard it sets is simultaneously a safety claim and a potential exhibit in the next cartel filing.
The labs that proposed SAFA are the same ones being sued for coordinating on safety. They are building the oversight mechanism while the legal system tests whether the oversight mechanism is the crime. The outcome of that test will determine whether SAFA becomes a governance milestone or a structural footnote in the era when safety coordination was legally possible.
Self-report caveat: The primary source for this story (The Information, September 23-24) is paywalled and was not directly retrieved. Chris Lehane’s OpenAI blog post is confirmed by search but returned HTTP 403 from Cloudflare. Dreamforce opposition confirmed via CNBC. Candidate leadership names confirmed via multiple secondary sources. BankInfoSecurity, CIO.com, Yahoo News, YourStory, ROIC.ai, and TheStreet corroborate the core details.