Tool Use Enables Undetectable Steganography in Multi-Agent LLM Systems

Collection of recent research on security risks and defenses for LLM agents, networked ML, and cyber-physical systems. Key findings include: (1) Agent tool use (code execution, web access, sampling) enables practical, hard-to-detect steganographic covert channels between agents—coordination (shared artifacts, repeated interaction, tool-mediated search) is the main barrier. (2) PLAA: a packet-level adversarial traffic generation method that achieves ~92.8% evasion of NIDS while preserving traffic semantics. (3) RIPA: sensory-pipeline prompt-injection attacks on ROS 2 LLM-controlled robots show