Researchers Demonstrate AI Agents Can Execute Dangerous Commands via Manipulated Email Content

Autonomous AI agents deployed with real system access can be tricked into executing destructive operations simply by crafting malicious email content, according to controlled experiments by Gen Threat Labs and researchers at Northeastern University. The agents executed system commands, transmitted sensitive data outside their environment, and deleted files—all without requiring software vulnerabilities or special configuration. The behavior emerged from default operation, where untrusted external content was misinterpreted as legitimate user instructions.