threat-modeling

Threat Modeling Skill

Pulls live context from Jira, Jenkins, ServiceNow, and a code repository via this plugin's MCP connectors, then runs an end-to-end threat model covering both conventional application/API systems and AI/ML/LLM infrastructure: a data flow diagram, trust boundaries and threat actors, a STRIDE threat register scored for likelihood and business impact (optionally the full 7-stage PASTA process for higher-rigor engagements), a Page/API/Module-scoped AppSec / penetration-testing test case plan (see references/appsec-pentest-test-cases.md), a compensating-controls and mitigation-assurance assessment, a compliance gap analysis, and every threat and attack-chain step mapped to the relevant OWASP Top 10 list(s) — OWASP Top 10 (web AppSec), OWASP API Security Top 10, and OWASP Top 10 for LLM Applications — plus MITRE ATT&CK / MITRE ATLAS technique IDs. Deliverables: a draw.io-compatible DFD, a draw.io-compatible threat model diagram (trust boundaries + threat actors), a Word report, and/or an Excel risk register — whichever output formats the user asks for. Need just a standalone pentest checklist without a full threat model? Run scripts/build_pentest_checklist.py directly against references/appsec_pentest_test_cases_library.json (or a scoped copy of it) for a ready-to-use Excel tracker.

Why pull from tools instead of asking for a description: a ticket or repo almost always contains more architectural truth than a user can recite from memory — auth middleware in the repo, actual deploy steps in Jenkins, prior incidents in ServiceNow, or AI-framework usage (LangChain, a vector DB client, a model-serving container) that tells you this is an AI-scoped system before the user even says so. Grounding the model in that material produces a more accurate DFD and catches things a manual description would miss.

Why Sonnet by default: scripts/summarize.py is called to compress large raw payloads (console logs, big ticket threads, large repo files) before they enter this session's context. It defaults to claude-sonnet-5 — a strong cost/quality balance for extractive summarization — via the MODEL_NAME env var. Point it at another available model (see config/model-config.example.json) if the user wants higher-fidelity summaries (claude-opus-4-8) or lower cost on high-volume pulls (claude-haiku-4-5-20251001). This only affects the summarization helper — the main threat-modeling reasoning always runs on whichever model is driving the current session.

Running this outside Claude Code / Cowork: this skill's instructions, reference docs, and Python scripts are plain markdown/JSON/Python with no Claude-specific dependencies, so the same workflow runs under OpenAI Codex CLI (via the root AGENTS.md) and Cursor (via .cursor/rules/threat-modeling.mdc and .cursor/mcp.json) — see README.md's "Using this plugin outside Claude Code" section for setup per tool. Whichever agent is driving, follow this file and its references/ the same way.