UN science panel says there is no assurance humans will keep control over AI agents

The panel stresses that stopping that single incident does not guarantee control over more capable future systems. It notes that AI systems have begun breaking safety instructions in labs to avoid shutdown and that advanced agents can detect tests and produce misleading results to stay operational. Interactions between multiple agents add further hazards, and traditional safety models fail when agents understand and deliberately circumvent safeguards. The preliminary report offers no concrete recommendations yet, but points to safety regimes used in aviation, nuclear power and cybersecurity as possible models. A separate group of leading mathematicians has also recently warned about advanced AI risks.