Skip to content

Security Policy (Offline-First)

Last Updated: 2026-07-14 Version: v0.2.0

Phase 4 CodeQL Security Resolution (2026-07-14)

Status: COMPLETE - All CodeQL security alerts resolved

Phase 4 has successfully eliminated all CodeQL security vulnerabilities through comprehensive remediation of untrusted code patterns in workflow_run privileged contexts. All 2 CRITICAL + 1 MEDIUM alerts have been definitively resolved.

Workflow_run Privileged Context Security Pattern

Key Security Principle: Do not use git operations in workflow_run contexts; use GitHub API validation instead.

Why This Matters: - CodeQL performs YAML-level dataflow analysis on workflow_run triggered jobs - Git operations (git fetch, git checkout) create untrusted code dataflow patterns - LGTM pragmas and code comments do NOT suppress workflow-level analysis - Only structural changes (removing git operations) eliminate the vulnerability

Recommended Pattern - Before (Vulnerable):

- name: Validate and sync
  run: |
    git fetch origin main --depth=1
    git checkout -fB _autogen_sync_ origin/main
    git add changes
    git push origin HEAD

Recommended Pattern - After (Secure):

- name: Validate and sync  
  run: |
    # Validate branch exists via API (no git fetch/checkout)
    gh api repos/${{ github.repository }}/branches/main --silent 2>/dev/null
    # (files already prepared, git operations only on trusted main branch)
    git add changes
    git push origin HEAD

Workflows Fixed in Phase 4: 1. iterative-self-healing-ci.yml - Removed 3 git fetch operations from workflow_run jobs 2. cognitive-analysis-feed.yml - Removed 2 git fetch + 2 git checkout operations 3. vars-guide-sync.yml - Removed 1 git fetch + 1 git checkout operation

For Complete Details: See CodeQL Alert Resolution Final Report


Reporting Security Issues

  • Please avoid committing secrets. Run python tools/security/scan_repo.py before PRs.
  • If a secret is suspected, rotate immediately and open a local remediation note; do not paste secrets into issues.

File Permissions Policy

Logging & Tracking Files

All log files and ML tracking artifacts use owner-only permissions (0o600) by default:

File Type Default Mode Rationale
NDJSON Logs 0o600 may contain API keys, connection strings
Tracking Summaries 0o600 Contains model hyperparameters, system metadata
Rotated Backups 0o600 Inherits sensitivity from active logs

Override for Shared Monitoring:

export CODEX_LOG_FILE_MODE=0o640  # Group-readable for monitoring agents

Security Note: Never use world-readable permissions (0o644) for production logs.

Implementation Details

The permission policy is enforced through: - src/codex_ml/logging/permissions.py - Central permission constants - src/codex_ml/logging/ndjson_logger.py - Structured log files - src/codex_ml/tracking/writers.py - ML tracking summaries

All os.open() calls use get_log_file_mode() to ensure consistent permissions across the codebase.

Security Scanning

# Run secret scanner
python tools/security/scan_repo.py

# Check license compliance
python tools/security/license_audit.py

# Snapshot dependencies
python tools/security/dep_snapshot.py

All scans are offline and output to audit_artifacts/ directory.

Dependency Management

  • Use requirements/lock.txt for reproducible installs
  • Run make deps to audit licenses and dependencies
  • Review audit_artifacts/license_audit.json periodically

Secret Handling

See docs/security/secret_handling.md for detailed guidance on: - Secret detection patterns - Incident response - Best practices