Detecting Shadow AI: Unsanctioned LLM Agents on Enterprise Endpoints
The paper introduces a taxonomy of nine classes of unsanctioned AI agents and six layers of endpoint footprint, arguing that process and connection structure alone cannot distinguish agents from non-AI behavioural twins. It presents Agent Detector, an open-source Windows detector that anchors behaviour to remote service identity and correlates configuration, installation and execution history using a catalog of 39 tools and 184 LLM endpoint entries. On a controlled bench of 12 synthetic agents, 3 AI components and 9 benign workloads repeated five times, two behavioural rules flagged every behavioural twin while Agent Detector flagged none and detected every agent in most runs. It reports temporal limits, including DNS-cache attribution lasting one to five minutes and intermittent agents seen in about one in three scans, at roughly 154 ms per scan.
It offers a concrete open-source detector and controlled evaluation for a growing enterprise security problem around unsanctioned LLM agents.
Comments & Questions (0)
No comments yet
Be the first to comment!