Neuro-Agentic Control: A Deep Learning-based LLM-Powered Agentic AI Framework for Controlling Security Controls
Introduces Neuro-Agentic Control, a framework coupling LLMs with Time-Series Foundation Models for physics-grounded autonomous defense in industrial IoT. Implements "Counterfactual Physics Injection" to simulate LLM interventions in the latent space, rejecting hallucinatory or unsafe actions before actuation. Achieves superior breach prevention rates (33.3%) compared to LSTM (26.7%) and TCN (13.3%) baselines on the SWaT dataset. Ensures zero physically invalid actions are executed, addressing th
Analysis
TL;DR
- Introduces Neuro-Agentic Control, a framework coupling LLMs with Time-Series Foundation Models for physics-grounded autonomous defense in industrial IoT.
- Implements "Counterfactual Physics Injection" to simulate LLM interventions in the latent space, rejecting hallucinatory or unsafe actions before actuation.
- Achieves superior breach prevention rates (33.3%) compared to LSTM (26.7%) and TCN (13.3%) baselines on the SWaT dataset.
- Ensures zero physically invalid actions are executed, addressing the safety liability of hallucinations in closed-loop control systems.
Why It Matters
This research addresses the critical safety gap in deploying generative AI for industrial control systems by providing a mechanism to validate LLM decisions against physical constraints. It offers a viable path for autonomous defense in critical infrastructure, where traditional rule-based systems fail and pure LLMs pose unacceptable risks due to hallucinations.
Technical Details
- Architecture: Combines an LLM-based planner (e.g., Gemini 2.5 Flash-Lite) for semantic reasoning with TimesFM, a pre-trained Time-Series Foundation Model, acting as a deterministic "Sentinel."
- Safety Mechanism: Utilizes "Counterfactual Physics Injection," which simulates proposed interventions within the foundation model's numerical latent space to predict outcomes and filter unsafe actions.
- Evaluation: Tested on the Secure Water Treatment (SWaT) dataset under stochastic attack scenarios, comparing performance against LSTM and TCN baselines.
- Performance Metrics: The framework prevented 33.3% of breaches below the threshold, outperforming LSTM and TCN, while maintaining a 100% rate of physically valid actions.
Industry Insight
- Hybrid Architectures for Safety: Critical infrastructure operators should consider hybrid models that pair high-level reasoning capabilities with deterministic, domain-specific foundation models to mitigate hallucination risks.
- Pre-Actuation Simulation: Implementing latent-space simulation layers can serve as a crucial safety gate for agentic AI systems, allowing for risk assessment before physical execution.
- Shift from Rule-Based to Adaptive Defense: As cyberattacks on OT become more sophisticated, moving beyond static rule-based monitoring to adaptive, physics-grounded AI frameworks is becoming essential for resilience.
Disclaimer: The above content is generated by AI and is for reference only.