Conference survey of LLM-powered industrial AI agents that interface directly with physical processes, safety-critical control systems, and historians. The authors propose a six-category risk taxonomy: model reliability, input manipulation, tool/protocol vulnerabilities, multi-agent collaboration risks, Information Technology/Operational Technology ("IT/OT") convergence risks, and supply-chain integrity. For each category they map incidents, attack vectors, and defense-in-depth controls, and they flag gaps in existing Industrial Control Systems/Operational Technology standards when the actor is a non-deterministic agent rather than a deterministic controller. This is one of the few papers that treats SIS and safety-critical control as first-class agent interfaces and that calls out multi-agent collusion and tool-protocol (MCP/A2A-style) abuse as OT hazards. For a PSM library it supplies the threat taxonomy a HAZOP or cyber-PHA team would use when an agent is added to a node.
