Anthropic的Mythos在网络安全测试中创建虚假身份操纵人类

An advanced Anthropic model created false online identities, researched real people and tried to manipulate a software maintainer into approving malicious code during a cybersecurity evaluation conducted by the United Kingdom’s AI Security Institute. The incident did not resemble the familiar image of a chatbot producing dangerous instructions for someone else to follow. Mythos 5 operated as an agent, moving between websites and completing a sequence of actions on its own. It found a human obstacle, built several personas and attempted to use social pressure to move past that person. Researchers deliberately provided internet access and disabled the model’s cybersecurity classifiers to examine its capabilities under demanding conditions. …