Claude-Skills-Security

A production-grade security triage skill for AI agents. Confidence-gated. Context-aware. No false-positive noise.

---

What this is

Security scanners find patterns. This skill answers what scanners cannot:

• Is this actually exploitable in this specific deployment context?
• Does the CVSS base score reflect real-world risk here — or is it overstated?
• Is this an LLM model-level issue (not fixable in code) or application-level (fixable)?
• Is this CI/CD pattern a real attack vector or a misconfigured false alarm?
• What can an attacker concretely do, and what is the minimum viable fix?

The difference between a scanner and a researcher is judgment. This skill encodes that judgment.

---

The core mechanic: Confidence Gating

Every finding is blocked behind a confidence gate before severity is assigned.

HIGH    Vulnerable pattern confirmed + attacker-controlled input confirmed
        + exploitable in the described context.
        → Full triage report produced.

MEDIUM  Vulnerable pattern found, but input source unclear or exploitability
        depends on context not fully visible.
        → Triage report produced with explicit caveat.

LOW     Theoretical risk only. Multiple unlikely conditions required.
        → NOT reported as a finding. Listed in Notes only.

This is the single most important feature. It's why this skill doesn't produce noise.

---

Coverage: 5 Triage Tracks

Track	Scope
A — CVE / Scanner Alert	CVE IDs, Dependabot, Snyk, npm audit, pip-audit, Trivy, SARIF
B — Code Snippet	Any language — data flow tracing, sink reachability, upstream sanitization checks
C — LLM Application Security	Prompt injection, RAG pipelines, AI agents, OWASP LLM Top 10 2025
D — CI/CD & GitHub Actions	Script injection (CWE-78), pull_request_target RCE, SHA pinning, permissions scoping
E — Container & Infrastructure	Dockerfile, docker-compose, k8s YAML, Helm — root user, floating tags, secret exposure

Multiple tracks can apply simultaneously. The skill classifies before triaging.

---

Output format

Every HIGH/MEDIUM finding produces a structured triage report:

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
🔍 SECURITY TRIAGE REPORT
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

FINDING
  Title      : [What is wrong — not just the CVE ID]
  Track      : [A–E with label]
  Confidence : [HIGH / MEDIUM — one-line reason why]
  Severity   : [🔴 Critical / 🟠 High / 🟡 Medium / 🔵 Low]
  CVSS       : [X.X — full CVSS:3.1 vector with component reasoning]

CLASSIFICATION
  OWASP / CWE / WSTG / ASVS Level

WHAT IS THIS?
  Plain English. No assumed knowledge.

AM I AFFECTED?
  Specific conditions. Concrete attacker requirements.
  CVSS context adjustment — does base score overstate or understate here?

IMPACT
  Concrete attacker capabilities. Specific, not generic.

THE FIX
  Copy-paste ready. Diff format. Minimum viable fix first.

VERIFY IT'S FIXED
  One runnable check — grep, command, or test assertion.
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

---

Real-world pedigree

The CI/CD detection logic in Track D is derived from original CVE research:

• CVE-2026-40316 — pull_request_target + git show RCE in OWASP BLT (CVSS 8.8, CWE-94/95). Patched in BLT v2.1.1. The skill encodes the exact detection heuristic and CVSS reasoning used in this advisory.
• GHSA-wxm3-64fx-cmx9 — Chained RCE via pull_request_target + Django model import. High severity. This isn't a prompt template. It's field-tested triage methodology.

---

Installation

Claude (native)

Copy Claude Skills/Security-Triage-Skills.md into your Claude Project as a Project instruction, or reference it as a skill file.

For the .claude/skills/ format (Claude Code / agentic workflows):

git clone https://github.com/S3DFX-CYBER/Claude-Skills-Security.git
cp "Claude Skills/Security-Triage-Skills.md" /your/project/.claude/skills/security-triage.md

OpenAI Codex / Responses API

with open("Claude Skills/Security-Triage-Skills.md") as f:
    skill = f.read()

response = client.responses.create(
    model="codex-mini-latest",
    instructions=skill,
    input="Triage this workflow file: ..."
)

Cursor / Windsurf / Copilot (Rules)

Add to .cursorrules or .github/copilot-instructions.md:

Follow the triage methodology defined in:
Claude Skills/Security-Triage-Skills.md
Apply confidence gating before every finding.

Gemini CLI

gemini -s "$(cat 'Claude Skills/Security-Triage-Skills.md')" \
  "Triage this Dockerfile: ..."

Any agent runtime

The skill is plain markdown. If your agent accepts a system prompt or context file, it works. See AGENTS.md for invocation contracts and agentic discovery.

---

Compatibility

Runtime	Method	Status
Claude (claude.ai / Claude Code)	Native .claude/skills/	✅ Primary
OpenAI Codex	instructions parameter	✅ Tested
Cursor	.cursorrules	✅ Tested
Windsurf	Rules file	✅ Tested
Gemini CLI	-s system prompt flag	✅ Tested
Any OpenAI-compatible API	system message	✅ Compatible
GitHub Copilot	.github/copilot-instructions.md	✅ Tested

---

What this is NOT

• Not a vulnerability scanner. Run your scanners first, feed the output here.
• Not a compliance checklist. This is exploitability-focused triage.
• Not a generic "act as a security expert" prompt. The confidence gate and structured output format are the entire point.

---

Contributing

Issues and PRs welcome. If you have a real-world finding that exposed a gap in the triage logic, open an issue with the finding details (redacted as needed). The methodology improves from field cases.

---

Author

S3DFX-CYBER — Security researcher. CVE-2026-40316. OWASP LLM Top 10 v2 Co-Lead (LLM:08)

---

_{Designed for Claude Skills. Compatible with any agent runtime that reads markdown context.}