Securing Encrypted Enterprise Networks: A Privacy-Preserving Approach to Detecting Autonomous AI Agents

This research presents a privacy-preserving approach for detecting autonomous AI agents in encrypted enterprise network traffic without decrypting packet payloads. The study investigates traffic side-channel characteristics, including packet sizes, inter-arrival times, and their temporal relationship, and introduces a novel Temporal-Sizing Burst Coherence (TSBC) metric. A Random Forest classifier is evaluated using engineered traffic features on synthetic network sessions, followed by a real-world passive traffic capture as a control. The work also proposes a passive out-of-band monitoring architecture suitable for integration with enterprise security monitoring systems. Future work considers adversarial traffic manipulation, sequence-based deep learning models, and high-throughput network evaluation.