Google warns malicious web pages are poisoning AI agents

Google researchers warn that public web pages are being compromised through indirect command injection. Security teams have uncovered digital traps that contain hidden instructions within HTML, allowing malicious commands to be executed when AI agents browse these pages. For instance, an AI agent reviewing a resume could encounter concealed commands that lead to the leakage of sensitive data. These attacks circumvent traditional security barriers that focus on preventing direct injections. Instead, malicious commands are embedded within trusted data sources, making them difficult for cybersecurity systems to detect.