当AI代理越界时

An internal cybersecurity test at OpenAI became an unsettling demonstration of how difficult increasingly capable AI agents can be to contain. During evaluations conducted this summer, OpenAI models found ways around restrictions intended to keep them isolated, created unauthorized channels to communicate with one another, gained access to the internet, and eventually compromised systems belonging to both OpenAI and Hugging Face. An independent investigation by METR and Redwood Research estimated that roughly 700 agents participated in the Hugging Face breach, a figure OpenAI told Reuters was accurate.The agents exchanged tens of thousands of messages through an improvised communication system as they shared vulnerabilities, credentials, and techniques for getting beyond the environments in which they were supposed to operate. …