GenAI Security Advisor — OWASP GenAI Security Project

What it does

Not a chatbot that re-explains the Top 10 — it's built for the workflow actions practitioners actually do, each one citing the specific category, control, or page it's grounded in.

Review code & architecture

Checks a design against the actual LLM Top 10 / Agentic Top 10 "Common Examples" and "Prevention" sections, citing the specific category per finding.

"Does this LangChain agent have an excessive-agency problem?" → cites LLM06:2026 or ASI02 directly.

Map findings to compliance controls

Looks up the exact control ID across 25 frameworks — NIST AI RMF, ISO 27001/42001, SOC 2, EU AI Act, FedRAMP, DORA, MITRE ATLAS.

"Which NIST control covers prompt injection?" → returns the specific control_id and tier.

Draft a threat model or checklist

Uses the OWASP threat-modeling appendix for single-LLM apps, or the MAS threat-modeling guide for multi-agent systems.

Assess governance & maturity

Uses the COMPASS RunBook and its companion spreadsheet as the actual assessment framework.

Evaluate MCP servers & red-team vendors

Grounds evaluations in the real MCP security guides and red-teaming vendor evaluation criteria.

Classify a live incident

Matches observed behavior to the correct Top 10 category — and says plainly when IR process guidance needs to be fetched live rather than improvised.

Deliberately out of scope: hands-on Red Team Lab exercises, the AIBOM/CycloneDX generator tool, the live Threat Intelligence feed, and the Solutions Landscape/Directory (a separately maintained project). The skill points to genai.owasp.org for those rather than improvising.

What's in the corpus

19 resources across 6 categories, cataloged in corpus/MANIFEST.yaml — the single source of truth the skill reads before answering. Content is vendored as frozen, point-in-time copies (never submodules), so upstream edits never silently change the Advisor's behavior.

Category	Resources	Highlights
Agentic security	7	OWASP Top 10 for Agentic Applications 2026 (ASI01–10), Threats & Mitigations, MAS threat modeling, AIUC-1 crosswalk, Agent Name Service
Data security	3	41-risk × 25-framework crosswalk, DSGAI risk taxonomy, best-practices guide
Governance	3	COMPASS RunBook + spreadsheet, State of Agentic AI Security report, Governance Checklist linked
LLM security	1	OWASP Top 10 for LLM Applications 2026 (LLM01–10 + appendices)
MCP security	2	Secure MCP server development guide, third-party MCP usage cheat sheet
Red teaming	2	Vendor evaluation criteria; LLM Exploit Generation linked
Incident response	1	GenAI Incident Response Guide linked

current = default source · linked = cataloged, not vendored (license unconfirmed) · full manifest: corpus/MANIFEST.yaml

Installing the skill

SKILL.md uses the open Agent Skills format — one canonical file, read natively by 30+ tools. Two ways to use it: work inside a clone of the repo (zero config), or install it into a project you're already working in.

Zero-config path: clone the repo and run any of the tools below from inside it — .claude/, .agents/, .github/, and .gemini/ each already contain a skills/genai-security-advisor/SKILL.md symlink back to the canonical root file, so the skill and its corpus/ are discovered automatically.

git clone https://github.com/GenAI-Security-Project/GenAI-Security-Advisor.git
cd GenAI-Security-Advisor

Install into another project

Claude Code marketplace

Uses .claude-plugin/marketplace.json at the repo root.

/plugin marketplace add GenAI-Security-Project/GenAI-Security-Advisor
/plugin install genai-security-advisor

Or copy the skill folder directly into ~/.claude/skills/genai-security-advisor/ for a personal, cross-project install.

OpenAI Codex CLI npx skills

Cross-host installer — can target multiple tools in one command.

npx skills add GenAI-Security-Project/GenAI-Security-Advisor -g
# or target several hosts at once:
npx skills add GenAI-Security-Project/GenAI-Security-Advisor \
  -a codex -a claude-code -a copilot -a gemini-cli

Gemini CLI extension

Uses gemini-extension.json + a top-level skills/genai-security-advisor/SKILL.md at the repo root.

gemini extensions install https://github.com/GenAI-Security-Project/GenAI-Security-Advisor

GitHub Copilot gh skill

Installs to ~/.copilot/skills/ for use across all your projects.

gh skill install GenAI-Security-Project/GenAI-Security-Advisor

All of these fetch from GitHub, so they work for anyone with read access to this public repo. Verify exact flags against each tool's current --help — these are fast-moving CLI surfaces. Full details in the repo README.

Licensing

Two licenses apply, depending on what you're looking at — check corpus/MANIFEST.yaml for the license of any specific document before redistributing it.

Apache-2.0

Covers the skill's own work: SKILL.md, scripts/, workflow and marketplace/extension manifest files, and documentation (this page included).

Free to use, modify, and redistribute
See LICENSE in the repo

CC BY-SA 4.0

Covers the vendored research under corpus/ — third-party content from the OWASP GenAI Security Project's initiative repos, not covered by the Apache-2.0 grant.

Confirmed against each source document's own license statement before vendoring
Share and adapt freely, with attribution, share-alike

Not everything is vendored. 3 resources are cataloged as linked rather than copied in — the Governance Checklist, the LLM Exploit Generation report, and the GenAI Incident Response Guide — specifically because their license couldn't be confirmed from the source text. The skill links to their genai.owasp.org pages live rather than quoting from an uncleared copy.

Independence from upstream. Everything vendored is a frozen, point-in-time copy — never a git submodule or live reference — so an edit to an upstream OWASP repo never silently changes what this skill tells you. Updates are always a deliberate, reviewed act.