Three companies that spend most of their time competing to ship the next frontier AI model faster than each other are now working together on something unusual: a body that would test and audit all of their models before any of them go public. Google, OpenAI and Anthropic are developing a proposed self-regulatory body, tentatively called the Standards Authority for Frontier AI (SAFA), that could launch by the end of 2026 or in early 2027, Emilia David reported for Bank Info Security on 24 September.
Where the idea came from
Google DeepMind chief executive Demis Hassabis proposed the concept in July, describing a standards body modelled in part on the Financial Industry Regulatory Authority (FINRA), the organisation that oversees US brokers and investment firms without being a government agency itself. Hassabis was direct about what building it would cost: “Funding would need to be substantial and likely mostly come from industry, in order to attract world-class technical talent.” Anthropic chief executive Dario Amodei added his own essay on the idea in September, and rival leaders responded within about an hour of each other online: Elon Musk endorsed the proposal, Sam Altman said OpenAI would commit to a first step toward it, and Hassabis tied the momentum back to his original July proposal.
The body’s proposed job is specific: developing common standards for assessing frontier AI models, running benchmark testing, and acting as an independent third party that can check a model before it reaches the public rather than after something has already gone wrong. Names already approached for leadership positions include Sriram Krishnan, a former White House AI policy adviser, Arati Prabhakar, a former Biden administration technology official, and former US Secretary of State Condoleezza Rice, among others, a roster that signals the companies want political credibility attached to the effort, not just technical expertise.
Not everyone is convinced
Cohere chief executive Aidan Gomez offered the sharpest public pushback: “AI needs guardrails… The dispute is over who writes them, who gets to participate and whose interests the rules are protecting.” That is the real tension in an industry writing its own rulebook: a body funded and staffed by the companies it oversees can credibly claim technical rigour while still facing a structural conflict of interest that a government regulator, whatever its own flaws, does not carry in quite the same form.
Why South African businesses should read past the Silicon Valley politics
The three companies behind SAFA are the same three whose models power a large share of the AI tools South African businesses already use daily, from customer service automation to the drafting and coding assistants inside enterprise software. Whatever safety and testing standards this body eventually sets will shape what those tools can and cannot do, and how much a business can rely on a vendor’s own safety claims versus needing to verify them independently.
The timing is also a pointed contrast with where South Africa’s own AI governance currently stands. Our coverage of the country’s own AI policy reset, delayed to March 2027 after its first draft was found to cite studies that do not exist, shows a government still working out the basics of the same problem three of the world’s best-resourced AI companies are now trying to solve for themselves, on their own timeline, funded by their own money. Whichever standard lands first, industry self-regulation or a national framework, is likely to shape the other.
