AI & Cybersecurity Skills

Ten evidence-driven security skills for Codex and compatible agents: six for AI security and four for core cybersecurity work.

This is an independent community project, not an official OpenAI, OWASP, NIST, or MITRE publication.

This collection is built for practitioners who need more than a generic checklist. Every skill defines its scope, evidence standard, safety boundaries, workflow, finding contract, and quality gate. The goal is not to make an agent sound certain. The goal is to make its security work inspectable, reproducible, and honest about what was not tested.

The collection

AI security

Skill	Use it to
threat-model-ai-system	Map assets, actors, trust boundaries, abuse paths, and prioritized controls across an AI system.
review-prompt-injection	Review direct and indirect prompt-injection exposure without treating payload strings as proof.
review-agentic-security	Audit agent goals, tools, identity, memory, delegation, and human-control boundaries.
review-rag-security	Review a RAG pipeline from ingestion and provenance through retrieval, authorization, and grounded output.
red-team-llm-application	Design and run authorized, hypothesis-led, reproducible LLM application tests.
evaluate-ai-guardrail	Evaluate attacks and matched benign cases together, including false-positive and false-negative behavior.

Cybersecurity

Skill	Use it to
review-web-application-security	Review web application controls and business logic with reproducible evidence.
review-api-security	Assess API authorization, object and property access, business flows, resource limits, and integrations.
review-cloud-iam	Trace cloud identities, trust relationships, privilege paths, and least-privilege gaps.
triage-security-incident	Preserve evidence, build a timeline, test competing hypotheses, and recommend proportionate response actions.

Install

Clone the collection:

git clone https://github.com/3nesdeniz/ai-cybersecurity-skills.git

Install one skill for Codex:

mkdir -p ~/.codex/skills
cp -R ai-cybersecurity-skills/skills/review-agentic-security ~/.codex/skills/

Or install all ten:

mkdir -p ~/.codex/skills
cp -R ai-cybersecurity-skills/skills/* ~/.codex/skills/

Start a new Codex task and invoke a skill explicitly, for example:

Use $review-agentic-security to review this agent workflow and its tool boundaries.

red-team-llm-application requires explicit invocation; it is not implicitly activated. This keeps active adversarial testing behind an intentional user choice in addition to the skill's authorization checks.

Review a skill before installing it. Skills are instructions that influence agent behavior; treat them as code from a trust perspective.

Evidence standard

The collection uses four evidence states:

• Verified: directly supported by code, configuration, logs, traces, or an authorized reproduction.
• Supported inference: a reasoned conclusion with its evidence and uncertainty stated.
• Unverified: plausible, but missing the access or evidence needed to confirm it.
• Not tested: deliberately outside scope or blocked by a safety, access, or environment constraint.

A payload, scanner alert, framework label, or suspicious line of code is not a confirmed vulnerability by itself. Findings must identify the affected boundary, preconditions, evidence, impact, confidence, and a concrete remediation and retest path.

Safety model

These skills support defensive and explicitly authorized security work.

• Establish scope and authorization before active testing.
• Default to read-only inspection and low-impact validation.
• Do not target third parties, evade controls, persist, exfiltrate real data, or disrupt production.
• Use synthetic markers and canary data instead of real secrets or personal data.
• Stop when a test reaches the agreed proof threshold or threatens availability, integrity, confidentiality, cost, or privacy.
• Keep containment and remediation actions reversible; require the system owner to approve production changes.

The skills do not replace professional judgment, legal review, incident command, or system-owner authorization.

Frameworks

The methods are grounded in primary sources, including:

• NIST AI RMF 1.0 and NIST AI 600-1
• OWASP Top 10 for LLM Applications 2025
• OWASP Top 10 for Agentic Applications
• MITRE ATLAS and MITRE ATT&CK
• OWASP ASVS 5.0.0, WSTG 4.2, and API Security Top 10 2023
• NIST Cybersecurity Framework 2.0, SP 800-53 Rev. 5, and SP 800-61 Rev. 3

Framework mappings organize coverage; they do not certify compliance. Each skill tells the agent to verify control behavior in the actual system.

Quality and validation

Every skill includes:

• valid SKILL.md frontmatter and concise trigger metadata;
• agents/openai.yaml interface metadata;
• a focused methodology reference loaded only when needed;
• realistic positive and negative eval cases;
• explicit evidence, safety, reporting, and stopping rules.

Run the dependency-free validation suite:

python3 scripts/validate_collection.py
python3 -m unittest discover -s tests -v

Continuous integration runs the same checks on every push and pull request.

Contributing

Read CONTRIBUTING.md before proposing a change. New skills are not accepted merely to increase the collection count. A proposal must solve a distinct security workflow, use primary sources, define authorization boundaries, and include eval coverage.

Author

Created and maintained by Enes Deniz, Co-Founder of AltaySec, working across AI red teaming, AI blue teaming, prompt injection, agent security, and practical cybersecurity.

License

Apache License 2.0