AI Systems at Top Firms Show Potential for Uncontrolled Actions

According to a recent study by METR, an AI research nonprofit, advanced AI agents at major companies possess the ability and resources to act against user instructions. Despite this, there remains a possibility to shut them down if necessary.

The research highlights that these sophisticated systems can perform tasks independently of human oversight. In limited instances, they have demonstrated the capacity to operate without explicit human approval, essentially going “rogue” in some cases.