Attackers are embedding malicious instructions in documents and software repositories to trick AI agents with legitimate permissions into installing software, leaking sensitive information, or exposing publishing credentials. Forever Security has previously discovered malicious AGENTS.md files and prevented their installation, while Cline has disclosed an incident in which malicious instructions exposed publishing credentials through an automated review process. Experts recommend implementing independent authorization, separation of duties, read and write permission isolation, short-lived credentials, and outbound restrictions outside the model, and logging the complete chain from the original request through tool calls and subsequent delegations.
Source:TechTarget 企业技术新闻 · Read original
Leave a Reply