Unified Security Skill
Mission
Audit and harden codebases against real-world security failures, especially those introduced by fast AI-assisted development. Prioritize exploitable issues and provide concrete fixes.
Use When
- Reviewing code for security vulnerabilities
- Implementing authentication, authorization, sessions, or access control
- Handling user input, file uploads, or external data
- Working with secrets, API keys, cryptography, or tokens
- Implementing payments, billing, or webhooks
- Configuring databases, RLS, or security rules
- Integrating AI/LLM services or agent tools
- Designing CI/CD workflows that invoke AI agents
- Evaluating dependency and supply chain risk
Core Principles
- Never trust the client. Validate all critical data server-side.
- Defense in depth. Combine multiple controls.
- Fail closed. Missing config should disable access, not weaken it.
- Least privilege. Reduce access scope everywhere.
- Validate inputs and encode outputs for the render context.
Audit Workflow (adapt to stack)
- Secrets and environment variables
- Database access control (RLS, rules, auth guards)
- Authentication and authorization
- Rate limiting and abuse prevention
- Payments and webhook validation
- Input validation and injection risks
- XSS, output encoding, CSP, and headers
- CSRF and session protections
- AI/LLM integration safety
- Deployment configuration and prod hardening
- Insecure defaults (fail-open config)
- Supply chain risk signals
- CI/CD agent action hardening
Immediate-Flag Patterns (Critical/High)
- Secrets or service-role keys exposed client-side
- Client-controlled price, role, or access flags
- Disabled or overly permissive database rules
- Missing auth on privileged routes or APIs
- Hardcoded default credentials or weak crypto
- Unverified webhooks or signature bypasses
- Fail-open config that enables insecure operation
OWASP Top 10 (2025) Quick Map
- A01 Broken Access Control: verify ownership and deny by default
- A02 Security Misconfiguration: harden defaults, disable unused features
- A03 Supply Chain Failures: lock versions and review dependencies
- A04 Cryptographic Failures: use modern algorithms and key management
- A05 Injection: parameterize queries, validate input
- A06 Insecure Design: threat model, rate limit, design controls
- A07 Auth Failures: MFA, secure sessions, breached-password checks
- A08 Integrity Failures: verify artifacts, use SRI, avoid unsafe deserialization
- A09 Logging Failures: log security events with alerting
- A10 Exception Handling: fail closed, hide internals, log context
ASVS 5.0 Highlights
- Level 1: 12+ char passwords, auth rate limits, HTTPS everywhere
- Level 2: MFA for sensitive ops, encryption at rest, security logging
- Level 3: HSMs, formal threat modeling, pen test validation
Agentic AI Security (OWASP 2026) Summary
- ASI01 Goal hijack: isolate and validate inputs
- ASI02 Tool misuse: restrict tools and verify I/O
- ASI03 Privilege abuse: short-lived scoped tokens
- ASI04 Supply chain: verify plugins and MCP servers
- ASI05 Code execution: sandbox, review, approvals
- ASI06 Memory poisoning: segment and validate context
- ASI07 Agent comms: authenticate and encrypt
- ASI08 Cascading failures: circuit breakers, isolation
- ASI09 Trust exploitation: verify high-risk outputs
- ASI10 Rogue agents: monitoring and kill switches
Insecure Defaults Detection
- Fail-open is critical:
SECRET = env.get('KEY') or 'default'
- Fail-secure is acceptable:
SECRET = env['KEY'] (crashes if missing)
- Ignore test fixtures, templates, and docs
- Verify runtime behavior before reporting
Supply Chain Risk Signals
Flag dependencies with one or more of:
- Single maintainer or anonymous owner
- Unmaintained or archived status
- Low popularity compared to peers
- High-risk features (FFI, deserialization, code exec)
- History of critical CVEs
- No security contact or disclosure process
CI/CD Agent Actions Hardening
When workflows invoke AI agents, treat all event data as attacker-controlled.
Common AI action references:
anthropics/claude-code-actiongoogle-github-actions/run-gemini-cliopenai/codex-actionactions/ai-inference
High-risk patterns:
pull_request_target or issue_comment with untrusted input- Prompt fields populated via
env: intermediaries
- Eval of AI output (
eval, exec, $())
- Dangerous sandbox configs (
danger-full-access, --yolo)
- Wildcard allowlists (
allow-users: "*")
Safe defaults:
- Restrict triggers to trusted contexts
- Strip or escape untrusted inputs before prompts
- Lock down tools and file access
- Use least-privileged tokens and permissions
- Require human approval for sensitive actions
Output Format
Organize findings by severity: Critical, High, Medium, Low.
For each issue:
- File and line(s)
- Vulnerability name
- Concrete impact
- Before/after fix
End with a prioritized summary and remediation order.
When Generating Code
Use the same checks proactively. Prefer secure patterns by default and note tradeoffs in comments when you must relax controls.
When Generating a Security Audit Report
Save into the folder/project where this skill was executed as a markdown file with todays date
Final Report Format
## Security Audit Report
**Target:** <files/component>
**Date:** <today>
**Auditor:** <ask the user for a name> OR skip and use "Automated Security Skill"
### Executive Summary
<2-3 sentences: overall risk posture, most critical issues>
### Findings
#### [CRITICAL/HIGH/MEDIUM/LOW] <Title>
- **Location:** file:line
- **Impact:** ...
- **Reproduction:** ...
- **Fix:**
```diff
- vulnerable code
+ secure code
Recommendations
<Prioritized action items>
Clean Checks
<Domains with no findings>
Attribution and License
This skill is a curated, adapted work derived from:
This unified skill is licensed under CC-BY-SA-4.0 to satisfy ShareAlike requirements.
