Experimental agentic ai cyber-defense system on a digital twin of an energy facility and a quantitative method for determining the permissible level of autonomy

Abstract

The increasing interconnection between information networks and control systems in energy facilities raises requirements for the speed of cyber-threat detection, while simultaneously creating a risk that automated responses may adversely affect the technological process. This article aims to develop a conceptual architecture for an agent-based cyber-defense system operating on a digital twin of an energy facility, along with a quantitative method for determining the permissible level of autonomy for defensive actions. The proposed architecture integrates telemetry collection and normalization, an asset graph, a monitoring agent, a planning agent, and a deterministic mechanism for policy verification and execution. The digital twin serves as a controlled, isolated environment for modeling system states and attack scenarios, as well as for validating decisions before they are applied to the operational network. The method accounts for asset criticality, the scale of potential cyber-physical consequences, upper bounds on errors (in detection, attribution, planning, and execution), telemetry quality, and the divergence between the digital twin and the physical system. For permitted actions, a dimensionless "risk pressure index" is calculated, which is then used to determine an autonomy level ranging from A0 to A4. Deterministic security constraints take precedence: commands prohibited by policy or absent from the allowlist are blocked, regardless of the index value; furthermore, autonomous execution is disallowed in the event of a critical loss of telemetry or an unidentified attack. The framework incorporates operator confirmation, strict separation between operational and test environments, hysteresis, and an audit log. The scientific novelty of the proposed approach lies in combining a formalized rule for delegating actions to an agent with the verification of safety constraints and the consideration of digital twin error—all within a unified policy framework. Illustrative scenarios and a plan for future validation are presented; the latter encompasses comparisons against manual response and rule-based systems, verification of safe action revocation, and uncertainty assessment. The article is conceptual and methodological in nature: the parameters and thresholds are illustrative, and no experimental results are claimed. Practical viability, safety, and the impact on response time remain to be established following the implementation of the digital twin and independent testing in an isolated environment.