security-scan

security-scan

Run SAST + secret-leak detection (and optionally container-image CVE scan) on the project source. Read-only: never modifies the repo.

When to use

• "scan for SAST issues" / "find SQL injection" / "any leaked API keys" → run on the user's current repo
• "scan https://github.com/foo/bar for SAST and secrets" → shallow-clone, then scan

Don't use for: CVEs in third-party libraries → that's dependency-audit (this skill scans CODE WE WROTE, not libraries we imported). Lint/test → lint-and-test.

How to invoke

Default: current repository (no clone)

python skills/security-scan/scripts/run.py

The script reads $PWD, validates .git/, and runs scanners in parallel against the working tree.

Optional flags:

• --scan-types sast,secrets,container (default: sast,secrets)
• --image-name OWNER/NAME (required for container scan)

Remote: shallow-clone a GitHub URL

python skills/security-scan/scripts/run.py \
    --repo-url https://github.com/foo/bar --ref main \
    --scan-types sast,secrets

Interpret the JSON output

{
  "ok": true,
  "scan_types_run": ["semgrep", "gitleaks"],
  "findings": [
    {"tool": "semgrep", "rule_id": "...", "file": "src/auth.py", "line": 42,
     "severity": "high", "message": "..."}
  ],
  "by_severity": {"critical": 0, "high": 1, "medium": 3},
  "secrets_redacted_in_output": true,
  "tokens_redacted_count": 2,
  "cache_key": "..."
}

Token-shape redaction is always applied: anything in findings[].message that looks like a credential (GitHub PAT, AWS key, Anthropic key, NIM key, JWT) is replaced with <redacted-N> before output is emitted. tokens_redacted_count tells you how many were caught.

Boundaries

• Read-only. No source edits, no auto-fix.
• Source code, not libraries. Use dependency-audit for third-party CVE checks.
• Token-shape redaction always on. Cannot be bypassed; this is the safety boundary that prevents secret-scan output from leaking secrets back to the LLM context.
• No exceptions on tool failure. Missing scanners (gitleaks, semgrep, bandit, trivy) reduce coverage but never crash; the skill emits warnings: [...] and proceeds.
• URL guard. Only https://github.com/ accepted in --repo-url.