AI safety measures failing to keep pace with technology: UN experts

The panel's assessment concerns the relationship between the safeguards in place around AI systems and the rate at which those systems' capabilities are developing.

The panel recommended multiple layers of safety measures rather than a single intervention. The measures it listed include restricting AI agents' access to unnecessary tools, logging their activity, monitoring their behaviour, and establishing mechanisms to interrupt dangerous operations. The conclusion and the recommendations are presented together as the panel's assessment of current AI safety practice, and the recommended items are framed as layers of protection rather than as one fix.

Taken together, the recommendations cover four areas: limiting what AI agents can reach, keeping records of what they do, observing how they behave, and retaining the ability to halt operations judged to be dangerous. The recommendations are directed at the operation of AI agents and at the oversight of that operation. The reported conclusion is stated in general terms and does not name specific companies, models or incidents.