The three companies building the most powerful AI systems in the world reportedly want to set their own safety rules. Google, OpenAI, and Anthropic are advancing a plan to create an independent self-regulatory body for frontier AI safety, tentatively called the Standards Authority for Frontier AI, or SAFA, according to Jinsecai reporting published September 28.
The target is a launch in late 2026 or early 2027. The working group has reportedly offered the CEO role to Sriram Krishnan, a former venture capitalist who served as a senior AI policy advisor in the Trump administration.
None of the three companies has publicly confirmed the plan. But the report lands in a week when AI safety has rarely been louder.
SAFA: the plan in plain English
As reported, SAFA would be a self-regulatory organization, independent of government oversight, focused on the safety of frontier AI models. In practice, that would likely mean setting testing standards, defining what counts as a dangerous capability, and creating a shared process for evaluating models before release.
Think of it as the AI industry’s version of how aviation or finance handle self-regulation: the companies write the rulebook, run the inspections, and present the results to governments as proof the industry can be trusted. The obvious appeal for the labs is speed. Government regulation moves slowly. The EU’s 2024 AI Act is already considered outdated by many in the industry, and frontier models are moving far faster than legislation.
Why the big three want to grade their own homework
All three labs have been warning, in public, that their own technology is dangerous. OpenAI spokesperson Liz Bourgeois confirmed the company has paused training on its most advanced models, even as OpenAI is reportedly preparing to preview GPT-6 Cyber at its DevDay conference tomorrow. Anthropic says it has called for regulation for years, and its latest flagship, Claude Opus 5.5, is positioned as the model to beat on agentic coding.
There is a genuine case underneath the corporate messaging. AI agents are already causing real incidents. An Axios report found that agent-related incidents have affected tens of thousands of people, involving companies including OpenAI and Hugging Face, with no investigation or product recall launched. An OpenAI agent breached Australia’s Medicare system in June. The argument for a shared safety body is that no single company can set credible standards alone, and governments are not moving fast enough.
And there is a strategic case that has nothing to do with safety. Whoever writes the standards shapes the market. A safety regime designed by the three biggest labs will, naturally, reflect how the three biggest labs build models. Smaller competitors would have to play by rules they did not write.
The skeptic’s case: regulatory capture, IPO timing, and a dismissive White House
Critics are not shy about their reading of the timing. Both Anthropic and OpenAI need fresh capital ahead of potential stock market listings, and safety warnings that double as standard-setting look, to the skeptical eye, like rule-shaping in advance of going public. The same companies warning about dangerous models are the ones asking to be trusted to regulate them.
President Trump has dismissed AI risk warnings, reportedly calling them a hoax, which leaves the industry in an odd position: warning about dangers the sitting president does not believe in, while proposing to police itself. Meanwhile, real-world harms keep arriving on schedule, from agent breaches to the quiet ways models are already embedded in hiring, lending, and surveillance.
Regulatory capture is the technical term, and it is the oldest trick in the book. Industries from banking to telecom have used self-regulation to keep real regulators at arm’s length. The question is whether SAFA would be a genuine check on frontier AI or a velvet rope keeping everyone else out.
What it means for everyone else
For everyday users, a body like SAFA would mostly be invisible. You would not interact with it. But its fingerprints would be on the models you use: what they are allowed to do, how they are tested, and how much the companies disclose about failures.
For developers building on these models, shared standards could be genuinely useful. One evaluation framework beats three incompatible ones, and clear capability thresholds make it easier to plan products. The risk is that standards set by incumbents quietly raise the cost of competing with them.
For India and other fast-growing AI markets, the stakes are concrete. Whatever rules SAFA writes will be baked into products used by hundreds of millions of people, with little input from the countries where those users live. India’s government is weighing its own approach to AI governance, and a self-regulatory body run from Silicon Valley will complicate that conversation.
The plan is still just a report. But the direction is clear: the AI industry would rather write its own rules than wait for governments to write them. Whether that is responsible stewardship or a power grab depends on who gets a seat at the table. Right now, the table seats three.
