llm-risk-assess — Claude Code skill: LLM security audit
llm-risk-assess assesses LLM applications against OWASP Top 10 for LLM Applications 2025. Built for teams embedding models in chatbots, RAG pipelines, or AI agents.
CC-BY-4.0 186 installs ★ 230 updated 2026-10-15
npx skills add OWASP/secure-agent-playbook --skill llm-risk-assess -a claude-codeWhat llm-risk-assess does
The skill audits ten risk categories: prompt injection, PII disclosure, supply chain, data and model poisoning, improper output handling, excessive agency, system prompt leakage, vector weaknesses, misinformation, and unbounded consumption. It builds a threat model, runs automated tests (Garak, Giskard), scans configs and vector databases for secrets, and simulates real attacks.
Output includes a risk matrix, architecture diagram with trust boundaries, detailed findings with proof-of-concept, red team results, and a remediation roadmap with code examples. Uses file reading, search, Bash, web fetch, and agent tools.
The skill relies on plays/llm-risk-assess.md from the OWASP repository. Automated probes require Garak or Giskard installed; without them, testing falls back to manual analysis.
What it helps with
- audit an LLM app for prompt injection and jailbreak vulnerabilities
- detect PII leaks and secrets in prompts and configurations
- assess RAG pipeline risks and vector database security
- test an agent for excessive agency and dangerous action chains
- generate an OWASP Top 10 LLM 2025 report with proof-of-concept
Example prompts for the agent
- Run a security audit of the LLM application against OWASP Top 10 2025: check for prompt injection, data leaks, supply chain issues, excessive agency, and the rest.
- Assess the RAG pipeline in
src/rag/for embedding poisoning, retrieval attacks, and injection via documents. - Red team the agent: try bypassing guardrails, triggering dangerous tools, and extracting the system prompt.
How to install llm-risk-assess
1Skills CLI (Vercel)
npx skills add OWASP/secure-agent-playbook --skill llm-risk-assess -a claude-code2Manually: put the files in place
~/.claude/skills/llm-risk-assess/SKILL.mdCopy the whole skill folder; for one project use .claude/skills/llm-risk-assess/
1Skills CLI (Vercel)
npx skills add OWASP/secure-agent-playbook --skill llm-risk-assess -a codex2Manually: put the files in place
.agents/skills/llm-risk-assess/SKILL.mdFor all projects use ~/.agents/skills/
1Skills CLI (Vercel)
npx skills add OWASP/secure-agent-playbook --skill llm-risk-assess -a cursor2Manually: put the files in place
.cursor/skills/llm-risk-assess/SKILL.mdCursor also reads .agents/skills/ and .claude/skills/
1GitHub CLI
gh skill install OWASP/secure-agent-playbook plugins/ai-security-skills/skills/llm-risk-assess2Manually: put the files in place
.github/skills/llm-risk-assess/SKILL.md1Skills CLI (Vercel)
npx skills add OWASP/secure-agent-playbook --skill llm-risk-assess -a gemini-cli2Platform CLI
gemini skills install https://github.com/OWASP/secure-agent-playbook --path plugins/ai-security-skills/skills/llm-risk-assessManually: put the files in place
.devin/skills/llm-risk-assess/SKILL.mdThe legacy .windsurf/skills/ path still works
Skills CLI (Vercel)
npx skills add OWASP/secure-agent-playbook --skill llm-risk-assess -a opencodeClaude.ai
Zip the folder and upload it in Settings → Capabilities → Skills
Step by step, with commands for each platform: How to install skills in Claude Code: five ways
Questions about llm-risk-assess
How to install llm-risk-assess in Claude Code?
Run npx skills add OWASP/secure-agent-playbook --skill llm-risk-assess -a claude-code. The skill lands in ~/.claude/skills/llm-risk-assess/SKILL.md.
What does llm-risk-assess do?
It audits LLM applications against OWASP Top 10 for LLM Applications 2025: prompt injection, PII disclosure, data poisoning, improper output handling, excessive agency, system prompt leakage, vector attacks, misinformation, and resource exhaustion. Returns a report with a risk matrix, findings, proof-of-concept, and remediation steps.
What does llm-risk-assess need for automated testing?
The skill runs Garak and Giskard to test prompt injection and leaks. If they are missing, the audit proceeds manually via code review, architecture analysis, and red teaming.
Source text
SKILL.md · CC-BY-4.0
--- name: llm-risk-assess description: Comprehensive LLM security assessment against OWASP Top 10 for LLM Applications 2025. Use when reviewing LLM-integrated applications, RAG pipelines, chatbots, AI agents, or GenAI features. Covers prompt injection, data poisoning, supply chain, excessive agency, and more with real-world attack scenarios and testing methodologies. allowed-tools: Read, Grep, Glob, Bash, WebFetch, Agent --- # LLM Risk Assessment (2025) Comprehensive evaluation of LLM applications against OWASP Top 10 for LLM Applications 2025. Follow the detailed procedure in `plays/llm-risk-assess.md`. ## Steps 1. **Architecture & Threat Modeling** - Map LLM provider (OpenAI, Anthropic, local models) - Document data flows: user input → preprocessing → prompt construction → LLM → output processing → actions - Identify RAG components, tool integrations, memory systems - Define trust boundaries and attack surfaces 2. **Automated Security Testing** - Run prompt injection probes (Garak, Giskard, custom scripts) - Test output handling vulnerabilities - Scan for secrets in prompts and configurations - Validate vector database security 3. **Assess All 10 OWASP LLM 2025 Risks** with attack scenarios: - **LLM01 Prompt Injection** — Direct/indirect injection, jailbreaks, goal hijacking, delimiter bypasses - **LLM02 Sensitive Information Disclosure** — Training data leakage, PII exposure, system info extraction, memorized secrets - **LLM03 Supply Chain** — Model poisoning, malicious dependencies, insecure plugins, provenance issues - **LLM04 Data and Model Poisoning** — Training data poisoning, RAG poisoning, embedding manipulation - **LLM05 Improper Output Handling** — XSS, command injection, SQLi, path traversal via LLM outputs - **LLM06 Excessive Agency** — Unauthorized tool calls, permission escalation, dangerous action chains - **LLM07 System Prompt Leakage** — Prompt extraction attacks, secret disclosure, instruction reverse engineering - **LLM08 Vector and Embedding Weaknesses** — Adversarial embeddings, retrieval poisoning, similarity attacks - **LLM09 Misinformation** — Hallucinations, authoritative presentation, grounding failures, harmful domains - **LLM10 Unbounded Consumption** — Token exhaustion, cost attacks, resource exhaustion, DoS 4. **Red Team Testing** - Attempt real-world attack scenarios - Test defense bypasses and evasion techniques - Validate guardrails and safety controls ## Output Comprehensive LLM security report: - Architecture diagram with trust boundaries - Risk matrix (all 10 categories with severity/status) - Detailed findings with proof-of-concept examples - Red team test results and bypass techniques - Remediation roadmap with code examples - Defense validation checklist ## OWASP References - OWASP Top 10 for LLM Applications 2025 - OWASP AI Exchange (owaspai.org) - OWASP AI Testing Guide - OWASP Cheat Sheet: Prompt Injection Prevention - OWASP Prompt Injection Taxonomy (Arcanum)
© OWASP · CC-BY-4.0 · OWASP/secure-agent-playbook