P2 Threat Model Generator Adds Agentic Interview Mode
GreenBladeSec updates its P2 threat modelling tool with an agentic interview mode to gather inputs through conversation instead of YAML files.
Stock photo for illustration only, not from the actual event
- P2 Threat Model Generator introduces an agentic interview mode.
- The LLM exclusively handles user conversation to populate ApplicationDescriptor.
- Core STRIDE rules and scoring engines remain fully deterministic.
- The tool runs on the in-house harness framework known as agent-core.
Developer GreenBladeSec has rolled out an update to the P2 Threat Model Generator, a Python-based utility designed to ingest configurations like Docker Compose, OpenAPI, and Kubernetes manifests, execute STRIDE analysis, score threats, and output HTML and JSON reports mapped to MITRE ATT&CK.
Originally launched back in June, the tool performed reliably through its deterministic rule engines, compliance matchers, and reporters. However, the initial friction point remained capturing the actual input data, as real-world threat modeling discussions rarely begin with a pre-written YAML file.
To solve this without compromising core reliability, every single deterministic component—including the STRIDE rule engine, CVSS-style scorer, compliance flag matcher, and HTML reporter—was kept completely untouched and driven by code.
Stock photo for illustration only, not from the actual event
Under this architecture, the LLM is restricted to a single responsibility: communicating with the user to figure out what parameters need to be fed into ApplicationDescriptor.
Consequently, the AI agent is explicitly not responsible for "generating threats" or performing security reasoning on its own. All the security logic remains strictly codified within the deterministic rule engine.
This architectural choice highlights a robust pattern for integrating AI into security tooling: isolating non-deterministic components solely to interface and data collection roles, while keeping compliance and scoring logic strictly deterministic to prevent hallucinations from compromising security assessments.
Both the earlier SAST/DAST triage tool (P1) and the updated P2 threat modelling tool now operate on top of the same unified in-house harness architecture named agent-core, streamlining how conversational interfaces are built and maintained across security utilities.
For further insights into software development and developer careers, platforms like DEV Community continue to serve as collaborative spaces powered by open-source technologies like Forem.
Source: Dev.to
Found something wrong in this article? Report an issue with this article
Comments
Leave a Comment