MATRA: Pragmatic Threat Modeling Framework for Agentic AI Systems

arXiv·high signal

MATRA provides a systematic method for practitioners to assess how known LLM threats translate into deployment-specific risks in agentic AI systems with tool access, databases, and external services. Validated via an OpenClaw case study, MATRA adapts established risk assessment methodology to the unique attack surfaces of agents — filling a gap where practitioners lacked structured threat analysis methods. Directly usable by teams deploying agents in production.