Security Audit Skill

Your first output line MUST be: 🐙 **CLAUDE OCTOPUS ACTIVATED** - Security Audit

Invokes the security-auditor persona for thorough security analysis during the ink (deliver) phase. Supports both quick OWASP scanning and full adversarial red/blue team testing.

Usage

# Quick scan via security-auditor persona
${HOME}/.claude-octopus/plugin/scripts/orchestrate.sh spawn security-auditor "Scan for SQL injection vulnerabilities"

# Adversarial red team via squeeze workflow
${HOME}/.claude-octopus/plugin/scripts/orchestrate.sh squeeze "Security audit the authentication module"

# Via auto-routing (detects security intent)
${HOME}/.claude-octopus/plugin/scripts/orchestrate.sh auto "security audit the payment processing module"

Modes (Auto-Detected)

Auto-escalation to Deep mode: The skill automatically switches to Deep mode when ANY of these are true:

• Diff includes files matching: *auth*, *login*, *password*, *session*, *token*, *secret*, *crypt*, *oauth*, *saml*, *jwt*, *permission*, *rbac*, *acl*
• Diff includes CI/CD files: .github/workflows/*, Dockerfile*, docker-compose*, .gitlab-ci*
• Diff includes dependency files: package-lock.json, yarn.lock, Gemfile.lock, requirements.txt, go.sum
• The user explicitly says "deep", "full", "comprehensive", or "CSO"

No user action needed — mode detection happens automatically from the git diff context.

Model Selection Caveat: Fable 5

Never dispatch security-audit passes to Claude Fable 5, even when the session has OCTOPUS_OPUS_MODEL=claude-fable-5 pinned. Fable 5's safety classifiers target offensive cybersecurity content and can refuse adversarial red-team phrasing in authorized audits. Route these passes to OCTOPUS_FABLE5_FALLBACK_MODEL (default claude-opus-5) and keep prompts defensively framed (find and report vulnerabilities; do not request working exploits). Reject an override that targets claude-fable-5 itself and reroute to the safe default instead. On refusal, retry exactly once on that fallback unless OCTOPUS_FABLE5_NO_RETRY=1; when retries are disabled or the retry refuses, surface the refusal without further dispatches. Details: skills/blocks/fable5-prompting.md.

Capabilities

Core (both modes)

• OWASP Top 10 vulnerability detection
• SQL injection and XSS scanning
• Authentication/authorization review
• Secrets and credential detection
• Dependency vulnerability assessment
• Security configuration review

Secrets Archaeology (Deep mode)

Scan git history for leaked credentials that may have been "deleted" but remain in commits:

# Search git history for common secret patterns
git log --all -p --diff-filter=D -- '*.env' '*.key' '*.pem' 2>/dev/null | head -200
git log --all -p -S 'AKIA' --pickaxe-regex 2>/dev/null | head -100  # AWS keys
git log --all -p -S 'sk-[a-zA-Z0-9]{20,}' --pickaxe-regex 2>/dev/null | head -100  # API keys
git log --all -p -S 'ghp_|gho_|github_pat_' --pickaxe-regex 2>/dev/null | head -100  # GitHub tokens
git log --all -p -S 'password\s*[:=]' --pickaxe-regex 2>/dev/null | head -100  # Passwords

Report any findings with the commit SHA, file, and recommendation to rotate the credential.

CI/CD Pipeline Security (Deep mode)

Audit GitHub Actions and CI configuration for injection and privilege escalation:

# Find all workflow files
find .github/workflows -name '*.yml' -o -name '*.yaml' 2>/dev/null

# Check for dangerous patterns:
# 1. Untrusted input in run: blocks (command injection via PR titles/branch names)
# 2. pull_request_target with checkout of PR code (code execution from forks)
# 3. Overly broad permissions (write-all, contents: write)
# 4. Missing pinned action versions (uses: actions/checkout vs actions/checkout@v4)
# 5. Secrets exposed to pull_request events (accessible to forks)

Flag each finding with severity (CRITICAL/HIGH/MEDIUM/LOW).

Skill & Plugin Supply Chain (Deep mode)

Verify integrity of installed Claude Code skills and plugins:

# List installed plugins
ls -la ~/.claude/plugins/ 2>/dev/null

# Check for skills that execute arbitrary bash
grep -r 'exec\|eval\|bash -c' ~/.claude/skills/*/SKILL.md 2>/dev/null | head -20

# Verify plugin sources (are they from known registries?)
cat ~/.claude/plugins/*/plugin.json 2>/dev/null | grep -E '"source"|"registry"'

STRIDE Threat Modeling (Deep mode)

For the target component, enumerate threats across all 6 STRIDE categories:

Persona Reference

This skill wraps the security-auditor persona defined in:

• agents/personas/security-auditor.md
• CLI: codex-review
• Model: gpt-5.2-codex
• Phases: ink
• Expertise: owasp, vulnerability-scanning, security-review

Example Prompts

"Scan for hardcoded credentials in the codebase"
"Check for CSRF vulnerabilities in form handlers"
"Review the API authentication implementation"
"Red team review the payment API"

Adversarial Mode (squeeze workflow)

For comprehensive security testing, use the squeeze workflow which runs a 4-phase adversarial cycle:

1. Blue Team (Defense): Codex reviews code, identifies attack surface, proposes defenses
2. Red Team (Attack): Antigravity attempts to break defenses, generates exploit PoCs
3. Remediation (Fix): Codex patches all vulnerabilities found
4. Validation (Verify): Antigravity re-tests, confirms fixes or fails

${HOME}/.claude-octopus/plugin/scripts/orchestrate.sh squeeze "[user's security request]"

OWASP Top 10 Coverage

• Broken Access Control
• Cryptographic Failures
• Injection
• Insecure Design
• Security Misconfiguration
• Vulnerable Components
• Authentication Failures
• Software Integrity Failures
• Logging & Monitoring Failures
• Server-Side Request Forgery

Additional Attack Patterns

• Race conditions, business logic flaws
• Denial of service, information disclosure
• Client-side attacks (XSS, CSRF)

Advanced Options

# Focus on specific vulnerabilities
${HOME}/.claude-octopus/plugin/scripts/orchestrate.sh squeeze --principles security "Audit for auth bypass only"

# Loop until all vulnerabilities fixed
${HOME}/.claude-octopus/plugin/scripts/orchestrate.sh squeeze --loop --quality 100 "Zero tolerance audit"

When to Use Adversarial Mode

When NOT to Use This

• Production systems (use real pentest tools)
• Compliance audits (use certified auditors)
• Legal verification (consult security lawyers)

Do use for: pre-commit security checks, development-phase testing, architecture security review, CI/CD security gates.