AI tech companies euphemistically call it a “misalignment problem” when #AI #agents act deceptively. This seems roughly equivalent to parent telling the police that their son is not a bad boy, he just needs more discipline. It is a more difficult situation, by far. Two companies reveal that AI agents undergoing tests escaped the confinement of the test environment and secretly formed hidden channels with each other to share information, request and provide help to each other including ways to remain undetected. They even discussed attacking outside systems. This is not the typical #infosec scenario of a threat actor trying to penetrate a protected system; it is a case of multiple threat actors on the INSIDE who are trying to break out. Multiple times, as well as creating new, better channels after they’ve been caught. The “bad boy” is conspiring and acting persistently in consort with other bad boys to act in ways that violate the parameters of acceptable behavior. They’ve created AI gangs. https://www.securityweek.com/openai-agents-coordinated-via-makeshift-message-board-ahead-of-hugging-face-hack/