← ClaudeAtlas

ai-threat-reviewlisted

Use when the user wants a defensive threat model or security review of an AI, LLM, RAG or agent system: STRIDE, OWASP Top 10 for LLM Applications, MITRE ATLAS, prompt injection and tool misuse.
ymurat/academic-agent-skills · ★ 0 · AI & Automation · score 72
Install: claude install-skill ymurat/academic-agent-skills
# AI threat review (defensive) Threat modeling and design review for systems that contain ML models, LLMs, retrieval or autonomous agents. Defensive only: produce threats, controls and test ideas, never working exploits or attack payloads against systems the user does not own. ## Sources in `resources/` - Upstream (Apache-2.0, credit: Inerthel, https://github.com/inerthel-agi/security-hardening): `resources/ai-threat-review/upstream/references/ai/` (llm-agent-security, mcp-security, rag-retrieval-security, prompt-and-tool-evidence-handling, agent-approval-patterns, multi-agent-boundaries-and-delegation, ai-agent-incident-response, agent-evals-red-teaming) and `.../threat-modeling.md`. License and NOTICE in `.../upstream/`. - OWASP LLM Top 10 (2025) list and ATLAS link: `resources/ai-threat-review/owasp-llm-2025.md`. Re-check the OWASP page for a newer edition before citing. ## Workflow 1. **Scope.** System purpose, users, assets (data, model weights, prompts, credentials, tools), trust boundaries, deployment. Draw a data flow diagram (Mermaid or `c4-model` container view) marking every boundary: user to app, app to model API, model to tools, retrieval store, external content. 2. **STRIDE per element.** For each process, store, flow and boundary: Spoofing, Tampering, Repudiation, Information disclosure, Denial of service, Elevation of privilege. Add AI-specific lenses: - untrusted content reaching the model (direct and indirect prompt injection via web pages, emails, fil