Reproduce “MATRA: Modeling the Attack Surface of Agentic AI Systems -- OpenClaw Case Study” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Tim Van hamme, Thomas Vissers, Javier Carnerero-Cano, Mario Fritz, Emil C. Lupu, Lieven Desmet, Dinil Mon Divakaran
MATRA introduces a pragmatic, impact-driven threat and risk modeling framework for agentic AI systems, systematically grounding generic LLM threats in specific deployed workflows. A case study with OpenClaw illustrates how the framework enables tracing attack paths and prioritizing mitigations through architectural controls against both adversarial and non-adversarial risks.
Key insights
- Prompt injection, being functionally analogous to code injection, cannot be fully eliminated; therefore, effective risk mitigation requires robust architectural controls rather than relying solely on input-level defenses.
- Autonomous tool execution in agentic systems introduces significant non-adversarial risks, where unexpected LLM behaviors (e.g., hallucination) can lead to high-impact outcomes even without malicious intent.
- Layered architectural defenses, including sandboxing, allow-listing, output sanitization, and least-privilege configurations, are critical for building robust security into agentic AI systems.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.