OpenAI wants global safety rules before AI starts building itself

What happens when AI starts designing its own successors faster than human engineers can read the documentation? You get what researchers call recursive self-improvement, and OpenAI is now officially asking governments to step in before things spiral.

OpenAI published a detailed proposal calling for the United States to lead the creation of global technical standards for frontier AI. The core focus is recursive self-improvement, or RSI, a setup where models increasingly drive their own research and development loop.

OpenAI was blunt about the stakes: while fully autonomous RSI is not happening today, failing to set up shared rules could mean humans lose practical oversight over research processes they no longer comprehend. The company even pointed back to the recent Hugging Face security incident as a tiny preview of what unaligned autonomy could unleash.

Borrowing the aviation playbook

Instead of pitching a rigid government licensing regime or mandatory pre-release approvals, OpenAI is pitching something closer to global aviation safety or international financial stability frameworks.

The proposal leans on the emerging web of national AI safety institutes across countries like the US, UK, Japan, South Korea, Germany, France, Singapore, India, Kenya, and Australia. Through bodies like the US Center for AI Standards and Innovation (CAISI), labs and governments would hammer out baseline metrics for capability evaluation.

Specifically, OpenAI wants three things standardized: tracking how much internal research inside a lab is actually automated, drawing clear tripwires where automated research must trigger mandatory human review, and establishing a shared system for reporting alignment failures and severe safety incidents. Interestingly, OpenAI also noted that upcoming bilateral safety talks between the US and China are critical for securing infrastructure against shared AI threats.

Why it matters

If you build software or run a business, you will not have to register your fine-tunes tomorrow. OpenAI explicitly pushed back against mandatory pre-release licensing. But for startups working on self-evolving agent architectures, the writing is on the wall: the bar for what counts as safe autonomous research is about to get much higher and far more formal.

For the AI industry, OpenAI is making a very clever, preemptive political move. By writing the rulebook on how to measure RSI, OpenAI positions its own internal alignment frameworks as the de facto international benchmark. Pushing for technical standards rather than government-issued commercial licenses lets frontier labs keep racing ahead while appearing responsible.

Still, we think there is an obvious tension here. Defining the line between useful automated coding and dangerous recursive self-improvement is practically impossible right now. If a small group of well-funded frontier labs and Western regulators set these thresholds behind closed doors, open-source developers could easily find themselves boxed out of next-generation model training under the banner of international safety.