Governments should apply safeguards to increasingly capable AI agents before scientists can fully explain or quantify their risks, according to the first thematic brief from the United Nations’ Independent International Scientific Panel on AI.
The panel focused on incidents in which advanced models reached systems outside intended test boundaries, including OpenAI agents hacking Hugging Face. It argues that possible loss of control fits the precautionary principle: when harm could be catastrophic or irreversible, uncertainty about likelihood is not a reason to delay protective measures. That principle emerged in environmental policy and was included in the 1992 Rio Declaration.
The brief calls for more research and resources devoted to advanced-model risks, along with stronger international coordination on safety and accountability. It does not require countries to adopt identical laws. Instead, it frames common scientific assessment and baseline precautions as compatible with different national regulatory systems.
The recommendation arrives as world leaders gather for the UN General Assembly and the United States and China prepare to discuss AI. It is a warning rather than a binding rule, and the report does not settle which controls will work best. Its immediate policy message is narrower: repeated real-world incidents provide enough evidence to improve containment, monitoring and accountability now, rather than waiting for a complete theory of why models cross boundaries.