MATRA: Modeling the attack surface for agentic AI systems, with OpenClaw as a case study

The researchers propose MATRA, a threat modeling framework for agentic AI systems, used to systematically evaluate which practical risks known LLM threats translate into in a specific deployment. The framework first conducts an asset-based impact assessment, then uses attack trees to determine the likelihood of these impacts occurring within that system architecture. The authors demonstrate it with a case study of a personal AI agent deployment built with OpenClaw, quantifying how architectural controls such as network sandboxing and least-privilege access reduce risk by limiting the scope of impact of successful injections.