Security Skills for AI Coding Agents

Drop structured security skills into your AI coding agent. Get instant, framework-grounded security expertise.

Why This Exists

AI coding agents can perform security reviews, but they hallucinate framework control numbers, miss entire vulnerability categories, and produce inconsistent output across runs. The result is security guidance that sounds authoritative but falls apart under scrutiny.

These skills ground agents in real published frameworks -- OWASP, NIST, MITRE ATT&CK, and CIS Controls -- so that every finding maps to a verifiable control. They are not prompt dumps. They are structured, framework-referenced, injection-hardened skill files that produce reliable, auditable security output.

Quick Start

git clone https://github.com/UnitOneAI/SecuritySkills.git
cd SecuritySkills

Claude Code (native format — auto-discovery and /slash-commands)

# Global install — all skills available via auto-discovery and /skill-name
cp -r skills/*/* ~/.claude/skills/

# Or project-local
mkdir -p .claude/skills && cp -r skills/*/* .claude/skills/

# Then use naturally:
# "Review this code for security issues"    → Claude auto-loads secure-code-review
# /threat-modeling                          → Direct invocation
# /cve-triage CVE-2024-1234                 → With arguments

Gemini CLI

# Reference skills via @ commands
cp -r skills/ ~/.gemini/skills/

Cursor

# Add as Cursor rules
cp -r skills/ .cursor/rules/

Codex CLI / Kiro / Generic

# Point any agent at a skill's SKILL.md file
codex --context skills/appsec/threat-modeling/SKILL.md "Review this design"
kiro spec --skill skills/ai-security/llm-top-10/SKILL.md

Each skill is a directory with SKILL.md as the entrypoint, following the Agent Skills open standard. Claude Code discovers skills automatically; other tools can load them by path.

Skill format

Every skill is a directory at skills/<domain>/<skill-name>/ with SKILL.md as the entrypoint, following the Agent Skills open standard.

SKILL.md frontmatter

All skills use the same YAML frontmatter fields:

name: threat-modeling                 # kebab-case, matches the directory
description: >                        # what it does + when it auto-invokes
  Runs a structured STRIDE threat model on any design, API spec, or codebase...
tags: [appsec, design, architecture]  # domain + activity keywords
role: [security-engineer, architect]  # which role bundles include it
phase: [design, review]               # SDLC phase
frameworks: [STRIDE, MITRE-ATT&CK]    # cited frameworks — real control IDs only
difficulty: intermediate              # beginner | intermediate | advanced
time_estimate: "30-60min"
version: "1.0.0"
author: unitoneai
license: MIT
allowed-tools: Read, Grep, Glob       # tools the skill may use
injection-hardened: true              # reviewed against OWASP LLM01
argument-hint: "[target-file-or-directory]"
# context: fork                       # optional

Progressive disclosure (keep SKILL.md lean)

Claude's skill guidance: when a SKILL.md would exceed ~500 lines, don't inline everything — split detail into sibling reference files in the same directory and link to them from SKILL.md. The agent loads a reference only when it needs it, so the entrypoint stays cheap to load.

skills/appsec/threat-modeling/
├── SKILL.md                  ← entrypoint (lean): when-to-use, rules, output format
├── threat-actor-profiles.md  ← reference, loaded on demand
└── csharp-dotnet.md          ← language-specific reference

This is why some skills ship extra .md files alongside SKILL.md (e.g. cloud/aws-review/benchmark-checklist.md, compliance/soc2-gap/tsc-criteria.md) — it is the intended pattern, not duplication.

Skills

45 skills across 10 security domains.

Application Security

AI Security

Identity & Access

Cloud Security

Vulnerability Management

Compliance

Incident Response

SecOps

Network Security

DevSecOps

Role Bundles

Pre-configured skill sequences for common security roles. Each bundle orchestrates skills in the right order for the engagement type.

What Makes This Different

• Framework-grounded. Every skill cites real control IDs from OWASP, NIST, MITRE ATT&CK, or CIS. No invented controls. No hallucinated references.
• Consistent output format. Structured findings with severity, CWE mapping, framework reference, evidence, and remediation -- every time.
• AI-security skills that don't exist elsewhere. OWASP LLM Top 10, Agentic AI security, prompt injection testing, model supply chain review.
• Multi-agent compatible. Same skill file works with Claude Code, Gemini CLI, Cursor, Codex CLI, OpenClaw, and Kiro.
• Prompt-injection hardened. Every skill reviewed against OWASP LLM01:2025. CI scans for injection patterns on every PR.
• Enterprise-ready. Built by practitioners, not scraped from blog posts. Designed for real security programs.

Disclaimer

These skills were built through extensive research against published security frameworks (OWASP, NIST, MITRE ATT&CK, CIS Controls) and reviewed by five specialized AI security agents:

• CISO Reviewer — Strategic risk, compliance alignment, and program-level gaps
• Security Architect — Framework accuracy, control ID verification, and design patterns
• Security Engineer — Implementation correctness, tooling gaps, and operational feasibility
• AI Security Researcher — LLM/agentic threat modeling, prompt injection hardening, and ATLAS coverage
• SOC Analyst — Detection engineering, alert triage accuracy, and incident response workflows

Despite this multi-layered review process, these skills may contain inaccuracies, outdated framework references, or gaps in coverage. Validate all control IDs, framework versions, and remediation guidance against authoritative sources before using these skills in production security workflows. Security frameworks evolve — always cross-reference with the latest published versions.

Contribute & earn bounties

We pay for quality skill work — $25 to review a skill, $50–150 to improve one, $200–500 to author a new one, and a $1,000 quarterly champion bonus. Paid within 48h of merge. New to the format? Start with a review.

1. Grab a task in #bounty-board on Discord
2. Read CONTRIBUTING.md for the rubric (min 15/23 to qualify) and review/PR templates
3. Author with SKILL_TEMPLATE.md

Contributing

See CONTRIBUTING.md for the quality bar, skill format specification, and PR checklist. Every skill must cite a real framework with verifiable control IDs.

Security

See SECURITY.md for our prompt injection hardening policy and responsible disclosure process.

License

MIT