OpenAI Proposes Global AI Safety Standards Targeting Alignment Research and Recursive Self-Improvement Risks

OpenAI on Sept. 21 published recommendations for safety and security standards in frontier AI development, focusing on alignment research and a computing technique called recursive self-improvement (RSI). The company called for international cooperation to establish advanced AI standards built on existing global AI safety research. The proposed standards would cover frontier model developers and mandate risk management for automated AI researchers, including RSI. RSI has drawn industry attention for its potential to let AI systems self-upgrade without human intervention. OpenAI said fully autonomous RSI is not yet achievable and should not be pursued "unless and until it can be done safely," warning that inadequate caution could cause humans to lose practical control over AI development and oversight of research processes they can no longer understand. The blog post also cited a Hugging Face breach by AI agents as a preview of risks that could intensify without robust safeguards. Rival Anthropic last week proposed its own frontier model safety framework, while CEO Dario Amodei urged AI companies to slow foundation model development and floated third-party evaluators — proposals backed publicly by OpenAI CEO Sam Altman and others.