Threat Modeling Software Development for LLM-Agent-Based Systems

Abstract

This article describes software that extends the capabilities of the OWASP Threat Dragon threat modeling project for LLM agent-based systems. New diagram components, attributes of new and existing components in Threat Dragon, and the ability to automatically identify threats based on rules are added. The structure defined by OWASP is taken as a reference structure of the LLM agent. The developed software adds support for LLM agent components, allows you to specify 27 new attributes, and automatically identifies threats based on 38 rules. A rule consists of a set of conditions under which a threat may occur, a description of a scenario taken from OWASP documentation, and the mitigations used for this threat, which can be proactive, reactive, and detective. The rule also specifies the threat type according to the STRIDE methodology. The flow diagram of the analyzed system is built manually, and the attributes of all components are also defined manually. Then, threats are automatically added to the diagram based on the diagram, attributes, and rules. If at least one condition is met, the threat will be added to the Threat Dragon diagram for the corresponding component. The authors defined a fairly simple syntax for the rules, which makes it easy to describe new threats, which will then be automatically added to the diagram. The development has received positive feedback from OWASP Threat Dragon leaders. It closes an important gap in the field of LLM agent security and offers a flexible, extensible tool for practical use. The article also provides an overview of seven threat modeling methodologies: STRIDE, LINDDUN, PASTA, STRIDE-AI, PLOT4ai, ADMIn and MAESTRO. An overview of threat modeling software is provided: OWASP Threat Dragon, OWASP pytm, Microsoft Threat Modeling Tool and ThreatFinderAI.