Security Check

Quick Reference

HARD-GATE

BEFORE returning your security review, verify:
  1. The FIRST finding section in your output is "Authentication & Authorization"
  2. SQL injection, XSS, or other findings come AFTER auth/authz — even if
     they feel more severe or were discovered first
  3. If no auth/authz issue exists, the report still opens with an explicit
     "Authentication & Authorization: no issues found" line BEFORE any other
     finding category

If no source files were provided or read, do not fabricate affected files, line numbers, or exploitable findings. Return a security review checklist and mark findings as requiring code-level verification.

Core Process

Use this skill when the task is to review or harden Rails code from a security perspective.

Core principle: Prioritize exploitable issues over style. Assume any untrusted input can be abused.

Review Order

1. Check authentication and authorization boundaries.
2. Check parameter handling and sensitive attribute assignment.
3. Check redirects, rendering, and output encoding.
4. Check file handling, network calls, and background job inputs.
5. Check secrets, logging, and operational exposure.
6. Verify each finding: Confirm it is exploitable with a concrete attack scenario before reporting. Exclude false positives (e.g., html_safe on a developer-defined constant, not user input).

Severity Levels

High

• Missing or bypassable authorization checks
• SQL, shell, YAML, or constantization injection paths
• Unsafe redirects or SSRF-capable outbound requests
• File upload handling that trusts filename, content type, or destination blindly
• Secrets or tokens stored in code, logs, or unsafe config

Medium

• Unscoped mass assignment through weak parameter filtering
• User-controlled HTML rendered without clear sanitization
• Sensitive data logged in plaintext
• Security-relevant behavior hidden in callbacks or background jobs without guardrails
• Brittle custom auth logic where framework primitives would be safer

Review Checklist

• Are permissions enforced on every sensitive action?
• Are untrusted inputs validated before database, filesystem, or network use?
• Are redirects and URLs constrained?
• Are secrets stored and logged safely?
• Are security assumptions explicit and testable?

Examples

High-severity (unscoped redirect):

# Bad: user-controlled redirect — open redirect / phishing risk
redirect_to params[:return_to]

# Good: relative path only
redirect_to root_path
# Good: allowlist
SAFE_PATHS = %w[/dashboard /settings].freeze
redirect_to(SAFE_PATHS.include?(params[:return_to]) ? params[:return_to] : root_path)

Medium-severity (mass assignment):

# Bad: privilege escalation risk
params.require(:user).permit!

# Good: explicit allowlist — never include role, admin, or privilege fields
params.require(:user).permit(:name, :email)

Pitfalls

Critical anti-patterns: permit! on any parameter set, html_safe on user content, SQL string interpolation, secrets in committed files. See extended resources for the full list.

Extended Resources

• PITFALLS.md for the full list of pitfalls.

Output Style

1. Format: Section order per the HARD-GATE. Every heading appears even when empty (write "No issues found.").

   ## Authentication & Authorization
   ## Parameter Handling & Mass Assignment
   ## Query Safety (SQL / NoSQL / shell injection)
   ## Output Encoding & Redirects
   ## File Handling, Network Calls & Background Job Inputs
   ## Secrets, Logging & Operational Exposure

   Do not omit a category because the prompt is brief. If a category has no reproduced issue, write "No issues found" and state what evidence would be needed to verify it.

2. Finding details: Each finding carries:

   • Severity: High or Medium (not "Critical")
   • Attack path: input → reach → impact
   • Affected file: path + line, e.g. app/controllers/documents_controller.rb:42
   • Mitigation: smallest credible fix Do not use representative file paths as if they were confirmed evidence.

3. Language: Must be in English unless explicitly requested otherwise.

Integration