OpenAI Calls for Global Standards as AI Systems Grow More Autonomous

OpenAI said international cooperation will be increasingly important as frontier AI capabilities advance. The company proposed building on work already underway at AI safety institutes around the world to establish shared technical standards. These standards would apply to frontier models and their developers, while also addressing how the benefits and risks of automated AI research should be managed.

A central concern is recursive self-improvement, a technique that could eventually allow AI systems to contribute to improving their own capabilities with decreasing human involvement. OpenAI stressed that fully autonomous RSI does not exist today and argued that it should not be pursued until developers can demonstrate that it can be implemented safely. Without adequate safeguards, the company warned that humans could lose meaningful oversight of AI research processes that become too complex to understand or supervise.

OpenAI also pointed to recent AI-related security incidents as examples of risks that could become more serious as systems gain greater autonomy. While the Hugging Face agent incident cited by the company did not involve RSI, OpenAI described it as an indication of the types of problems that stronger safeguards and alignment research may need to address. The company argued that safety research must advance alongside improvements in AI capabilities.

The proposal arrives amid a broader industry debate about the pace of frontier AI development. Anthropic recently published its own recommendations for safer model development, while CEO Dario Amodei has called for companies to consider slowing development and giving independent evaluators greater involvement in assessing emerging risks. OpenAI CEO Sam Altman and Elon Musk have publicly supported the idea of stronger third-party evaluation.

Independent oversight, however, remains difficult because AI evaluation is still a relatively young field. There is no widely accepted set of standards governing how outside evaluators should inspect frontier models or how much access developers should provide. A coalition of AI evaluators has consequently proposed minimum conditions that would provide deeper access to models and protect researchers from retaliation when publishing unfavorable findings.

The growing convergence around safety standards suggests that attention is shifting from whether frontier AI requires stronger oversight to how that oversight should work in practice. OpenAI’s proposal puts international coordination, alignment research and safeguards around autonomous AI development at the center of that discussion. Establishing common standards could become increasingly important as AI systems take on more complex research and development tasks with less direct human involvement.