dyoshikawa/security-scan-diff

target_ref = $ARGUMENTS

If target_ref is not provided, ask the user which tag or commit to compare against HEAD.

Overview

Thoroughly check for malicious code in the diff between ${target_ref} and the latest commit (HEAD).

Steps

1. Verify the target ref exists and get the diff scope.

   • Run git log ${target_ref}..HEAD --oneline to list commits.
   • Run git diff ${target_ref}..HEAD --stat to get file change statistics.
   • Categorize changed files into: CI/CD workflows, source code, and config/docs.

2. Execute the following security reviews in parallel using subagents:

   • Call security-reviewer subagent to review CI/CD and workflow files (.github/, scripts/) for:

     • Secret exfiltration
     • Script injection (${{ github.event.* }} direct expansion in run:)
     • Suspicious external URLs/API connections
     • Privilege escalation or token misuse
     • Malicious command execution (curl | bash, eval, base64 decode execution)
     • Supply chain attack patterns (suspicious npm packages, unsigned action references)
     • Dangerous pull_request_target usage

   • Call security-reviewer subagent to review source code files (src/) for:

     • Arbitrary code execution (eval, Function constructor, suspicious child_process usage)
     • Path traversal (../.. directory escape)
     • Command injection (user input passed directly to shell commands)
     • Suspicious external communication (fetch, http.request, axios to external URLs)
     • Unauthorized filesystem operations
     • Credential/token leakage (hardcoded tokens, logging sensitive values)
     • Dependency tampering (suspicious package.json changes)
     • Backdoor patterns (obfuscated code, suspicious conditionals, hidden functionality)
     • Prototype pollution and deserialization vulnerabilities
     • Supply chain attacks (suspicious new dependency packages)

   • Call security-reviewer subagent to review config and documentation files for:

     • Suspicious dependencies or scripts in package.json
     • Suspicious registries or URLs in lockfiles
     • Security rule relaxation in config schemas or linter configs
     • Suspicious settings in devcontainer or editor configs
     • Phishing URLs in documentation
     • Malicious instructions in AI rule/subagent/skill definitions

3. Integrate the results from all subagents and produce a unified report in the following format:

   ## Security Review Report: ${target_ref} -> HEAD
   
   ### Conclusion
   - Whether malicious code was detected or not
   
   ### Check Results Summary Table
   | Check Item | Result |
   |------------|--------|
   | ... | ... |
   
   ### Findings (if any)
   | Severity | Description | File | Risk |
   |----------|-------------|------|------|
   | ... | ... | ... | ... |
   
   ### Recommendations (if any)
   - Actionable recommendations for each finding
   
   ### Positive Observations
   - Good security practices found in the diff