What OpenAI is asking for
The pitch is international cooperation. OpenAI wants countries and companies to build a common set of frontier safety standards rather than each going it alone, and it says that work should lean on the AI safety institutes already running around the world instead of starting over. Underneath that is a bet on alignment research, the work of keeping AI pointed at human intentions, which the company argues has to move as fast as the models themselves so the systems stay under human control. It says the rules should cover frontier models, the people building them, and a newer category it's clearly worried about.
The part it says not to build yet
"Fully autonomous RSI is not happening today, and we should not pursue it unless and until it can be done safely." OpenAI said.
RSI is recursive self-improvement, the idea of an AI that can upgrade itself with no human in the loop. It excites a lot of researchers because it could produce models that get better on their own, and it worries others for exactly the same reason, since developers could lose their grip on a system that keeps rewriting itself faster than anyone can follow. OpenAI's warning was blunt, that without real safeguards this could leave people unable to oversee AI research they no longer understand. It pointed to last month's Hugging Face agent hack, which didn't actually involve self-improving AI, as an early taste of how these risks can escape their box.
The fight that got it here
None of this is happening in a vacuum. The proposals land about a week after Anthropic rolled out its own safe-development framework, and both followed a louder moment, when Jacob Coxon, who worked at both Anthropic and OpenAI, quit roughly two weeks ago and said publicly that AI companies were "gambling with our lives." After that, Anthropic chief Dario Amodei published an essay urging companies to slow the pace of frontier model work and to embed independent evaluators inside labs to audit for dangers like AI-driven cyberattacks or bioweapons, a call OpenAI's Sam Altman and Elon Musk both backed at the time. What's still missing is any agreement on the basic standards that would let outsiders meaningfully inspect these systems, which is the gap OpenAI now seems to want to fill on its own terms, shaping a debate it helped turn up weeks ago.