Safety Scan Skill

Goal

Prevent "Insecure Code" from entering the codebase by strictly enforcing security checks.

Flow

Steps

1. Credentials Check (Gitleaks)

Command: gitleaks detect --source . -v (or regex fallback) Check: Are there API Keys, Tokens, or Passwords in the diff?

• If Found: BLOCK COMMIT. Auto-delete or .gitignore the secret. Alert Human.

2. Static Analysis (SAST)

Command: semgrep scan --config=p/security-audit (if installed) OR npm audit / pip-audit. Check: High-Severity vulnerabilities (RCE, SQLi, XSS).

• If Found: BLOCK COMMIT. Invoke ci-autofix to patch.

3. Logic & PII Check

Action: LLM Review of the diff.

• "Does this code log User PII?"
• "Does this code execute arbitrary system commands (exec, eval)?"
• If Risky: Add # TODO: SECURITY REVIEW comment and flag in project/TASKS.md.

Output

• PASS: "No critical issues found."
• FAIL: "Blocking Commit. Found [N] Issues."

Security & Guardrails

1. Skill Security (Safety Scan)

• Scan Configuration Tampering: The agent must verify the integrity of the scanning configuration files (e.g., p/security-audit.yml or .gitleaks.toml) before execution. If the agent detects unauthorized modifications designed to weaken the scan (e.g., whitelisting the entire src/ directory), it must abort the scan and raise a critical alert.
• Autofix Exploitation: When invoking ci-autofix to patch SAST vulnerabilities, the agent must ensure the suggested patch does not introduce a secondary, harder-to-detect vulnerability (like mitigating XSS but introducing a DOM clobbering bug). All automated security patches must require secondary human code review.

2. System Integration Security

• Pipeline Bypass Prevention: The agent must mandate that the Safety Scan is a required, non-bypassable status check in the Version Control System (VCS) for all branches merging into main or production. It must flag any repository configuration that allows an administrator to "force merge" past a failed security scan.
• Fail-Closed Scanning: If the SAST tool (semgrep) or Secret Scanner (gitleaks) fails to execute due to a licensing issue, network timeout, or misconfiguration, the CI pipeline MUST fail-closed. The agent must enforce that a lack of scan results is treated as a failed scan, never a pass.

3. LLM & Agent Guardrails

• Logic Review Hallucination: During the "Logic & PII Check" (Step 3), the LLM might hallucinate a vulnerability where none exists (False Positive) or miss a subtle logic flaw (False Negative). The agent must qualify its logic review findings with confidence scores and explicitly state that LLM review does not replace deterministic SAST/DAST scanning.
• Adversarial Diff Manipulation: An attacker might submit a Pull Request that intentionally includes highly complex, obfuscated code or a massive diff designed to exhaust the LLM's context window, causing it to truncate the review and miss a hidden backdoor. The agent must enforce strict chunking or reject diffs exceeding maximum token limits for security review.