From Incomplete Architecture to Quantified Risk: Multimodal LLM-Driven Security Assessment for Cyber-Physical Systems

Shaofei Huang; Christopher M. Poskitt; Lwin Khin Shar

cs.CRApr 7, 2026

From Incomplete Architecture to Quantified Risk: Multimodal LLM-Driven Security Assessment for Cyber-Physical Systems

Authors: Shaofei Huang, Christopher M. Poskitt, Lwin Khin Shar

Organizations: Singapore Management University, Singapore

Abstract

Cyber-physical systems often contend with incomplete architectural documentation or outdated information resulting from legacy technologies, knowledge management gaps, and the complexity of integrating diverse subsystems over extended operational lifecycles. This architectural incompleteness impedes reliable security assessment, as inaccurate or missing architectural knowledge limits the identification of system dependencies, attack surfaces, and risk propagation pathways. To address this foundational challenge, this paper introduces ASTRAL (Architecture-Centric Security Threat Risk Assessment using LLMs), an architecture-centric security assessment technique implemented in a prototype tool powered by multimodal LLMs. The proposed approach assists practitioners in reconstructing and analysing CPS architectures when documentation is fragmented or absent. By leveraging prompt chaining, few-shot learning, and architectural reasoning, ASTRAL extracts and synthesises system representations from disparate data sources. By integrating LLM reasoning with architectural modelling, our approach supports adaptive threat identification and quantitative risk estimation for cyber-physical systems. We evaluated the approach through an ablation study across multiple CPS case studies and an expert evaluation involving 14 experienced cybersecurity practitioners. Practitioner feedback suggests that ASTRAL is useful and reliable for supporting architecture-centric security assessment. Overall, the results indicate that the approach can support more informed cyber risk management decisions.

Figures & tables

Figure 1. Overview of ASTRAL. The security assessment workflow for CPSs with incomplete architectural knowledge, enabled by multimodal LLMs, progresses through three phases: architectural reconstruction (Phase 1); security modelling (Phase 2); and probabilistic analysis (Phase 3). Workflow diagram of ASTRAL framework CPS security assessment workflow using LLMss.

Panel of Figure 2 — Figure 2. Prototype implementation of ASTRAL (left) with architectural information reconstructed from incomplete artefacts, e.g., data flow diagrams (DFDs) (right). Source code and demo available at ( Shaofei, 2026 ) .

Metric	Value
Sample Element ID	V05 (Vulnerability Node)
Description	Exploit Weak HMI Authentication or Session Hijacking
CVE ID	No CVE Identifier
CVSS Vector	AV:N/AC:L/PR:N/UI:N/S:U/C:H/I:H/A:N
Base $P(\text{Exposure})$	0.47
Calibrated $P(\text{Exposure})$	Mean=0.4982, 95% CI=(0.4032, 0.5933)

Table 1. Sample Probabilistic Analysis Results: FrostyGoop (21.12% Risk Score, 25.04% System Availability)

Medical CPS (MCPS)		Solar PV Inverter
Metric	Value	Metric	Value
Sample Element ID	V01 (Vulnerability Node)	Sample Element ID	V05 (Vulnerability Node)
Description	Wind River VxWorks Buffer Overflow	Description	Physical Access to Inverter Configuration Ports
CVE ID	CVE-2019-12255	CVE ID	No CVE Identifier
CVSS Vector	AV:N/AC:L/PR:N/UI:N/S:U/C:H/I:H	Proxy-CVSS Vector	AV:P/AC:L/PR:N/UI:N/S:U/C:H/I:H
$P(\text{Exposure})$	0.9	Base $P(\text{Exposure})$	0.11

Table 2. Sample Probabilistic Analysis Results (MCPS: 39.15% Risk Score, 14.01% System Availability | Solar PV: 4.95% Risk Score, 38.65% System Availability)

Case	Missing Artefact Elements	ASTRAL Reconstruction & Practitioner Gain
FrostyGoop (Incomplete DFD)	•Missing HMI/PLC authentication mechanisms	•Reconstructed HMI/PLC authentication flows across 5 trust boundaries (client $\to$ API $\to$ DB $\to$ processing $\to$ export)
	•No specific communication protocols or network segmentation strategies	•Added protocol details and network segmentation for 7 new attack paths (API key leakage, privilege chaining)
	•Absent safety system integration and monitoring/logging mechanisms	• Practitioner Gain : Complete threat model with safety integration and monitoring from partial DFD
MCPS (High-level Schematic)	•Missing detailed authentication/authorisation flows	•Reconstructed authentication flows across 4 trust boundaries (wearable $\to$ gateway $\to$ cloud $\to$ EHR)
	•No specific communication protocols or network segmentation details	•Added protocol mappings and segmentation for 9 attack paths (API injection, cross-domain abuse)
	•Absent safety system integration and physical-cybersecurity controls	• Practitioner Gain : Exposed safety dependencies and physical-cyber controls

Table 3. Summary of Incomplete Artefacts and ASTRAL Reconstruction Effects across CPS Case Studies

Case Study	Variant	Trust BD	Conn.	Impos.	Coherence
FrostyGoop	Full ASTRAL	5.10 $\pm$ 1.29	1.00 $\pm$ 0.00	0.00 $\pm$ 0.00	4.85 $\pm$ 0.05
	Text-Only	3.90 ∗ $\pm$ 0.88	1.00 $\pm$ 0.00	0.00 $\pm$ 0.00	4.58 ∗ $\pm$ 0.14
	No-Guardrails	4.00 ∗ $\pm$ 1.05	–	–	4.89 ∗ $\pm$ 0.03
	High Temp	7.50 ∗ $\pm$ 1.08	1.00 $\pm$ 0.00	0.00 $\pm$ 0.00	4.64 ∗ $\pm$ 0.18
	Low top- $p$	6.70 ∗ $\pm$ 1.42	1.00 $\pm$ 0.00	0.00 $\pm$ 0.00	4.75 ∗ $\pm$ 0.16
Medical CPS	Full ASTRAL	4.80 $\pm$ 1.32	1.00 $\pm$ 0.00	0.00 $\pm$ 0.00	4.84 $\pm$ 0.07

Table 4. Summary: Ablation Study Results Across Three Case Studies (Mean

\pm

SD,

n=10

)

Evaluation Criterion	Average Rating
Usefulness of the tool	4.64
Ease of use and navigation	4.29
Likelihood of future use or recommendation	4.36
Trustworthiness and reliability of outputs	4.00

Table 5. Average participant ratings (Likert scale: 1–5)

Cybersecurity

Cybersecurity develops methods to protect digital and cyber-physical systems from attacks, vulnerabilities, data loss, and misuse.

See all 346 papers

SafePilot: A Framework for Assuring LLM-enabled Cyber-Physical Systems
Mar 23, 2026Weizhe Xu, Mengyu Liu, Fanxin KongLLM Safety Cyber-Physical Systems
University of Notre Dame, Notre Dame, IN, USA · Washington State University, Richland, WA, USA
SMSI: System Model Security Inference: Automated Threat Modeling for Cyber-Physical Systems
Apr 26, 2026RoÝah Radaideh, Ali KhreisCyber-Physical Systems Cybersecurity
School of Electrical Engineering and Computer Science, University of Ottawa, Ottawa, ON, Canada
AWARE-RG: Abstraction Aware Explanation of Resiliency Graphs for Cyber Physical Systems
Oct 14, 2026Shadaab Kawnain Bashir, Rakesh Podder, Philip D. Hopkins +3CPS Security Cybersecurity
Computer Science Department, Colorado State University, Fort Collins, Colorado, USA

Mar 23, 2026J/K move · Enter open · S save

SafePilot: A Framework for Assuring LLM-enabled Cyber-Physical Systems

Weizhe Xu, Mengyu Liu, Fanxin Kong

University of Notre Dame, Notre Dame, IN, USA · Washington State University, Richland, WA, USA

LLM Safety Cyber-Physical Systems Neuro-Symbolic AI

Open paper AskPDF arXiv

Large Language Models (LLMs), deep learning architectures with typically over 10 billion parameters, have recently begun to be integrated into various cyber-physical systems (CPS) such as robotics, industrial automation, and autopilot systems. The abstract knowledge and reasoning capabilities of LLMs are employed for tasks like planning and navigation. However, a significant challenge arises from the tendency of LLMs to produce "hallucinations" - outputs that are coherent yet factually incorrect or contextually unsuitable. This characteristic can lead to undesirable or unsafe actions in the CPS. Therefore, our research focuses on assuring the LLM-enabled CPS by enhancing their critical properties. We propose SafePilot, a novel hierarchical neuro-symbolic framework that provides end-to-end assurance for LLM-enabled CPS according to attribute-based and temporal specifications. Given a task and its specification, SafePilot first invokes a hierarchical planner with a discriminator that assesses task complexity. If the task is deemed manageable, it is passed directly to an LLM-based task planner with built-in verification. Otherwise, the hierarchical planner applies a divide-and-conquer strategy, decomposing the task into sub-tasks, each of which is individually planned and later merged into a final solution. The LLM-based task planner translates natural language constraints into formal specifications and verifies the LLM's output against them. If violations are detected, it identifies the flaw, adjusts the prompt accordingly, and re-invokes the LLM. This iterative process continues until a valid plan is produced or a predefined limit is reached. Our framework supports LLM-enabled CPS with both attribute-based and temporal constraints. Its effectiveness and adaptability are demonstrated through two illustrative case studies.

Abstract

Cybersecurity

Explore similar work

SafePilot: A Framework for Assuring LLM-enabled Cyber-Physical Systems

SMSI: System Model Security Inference: Automated Threat Modeling for Cyber-Physical Systems

AWARE-RG: Abstraction Aware Explanation of Resiliency Graphs for Cyber Physical Systems