langdoctor 🩺

Scan your LangGraph/LangChain project for known CVEs, insecure configs, and production footguns — in seconds, offline, no API key.

brew doctor for your agent stack.

Quickstart

pipx run langdoctor        # or: uvx langdoctor

Run it in your project directory. langdoctor reads your dependency files and source, reports issues grouped by severity, and exits non-zero when it finds something at or above your threshold.

langdoctor                 # scan the current directory
langdoctor path/to/project # scan a specific directory
langdoctor list-checks     # every check, with IDs and severities
langdoctor --version       # tool version + advisory DB date

Why

2026 has been a rough year for the agent stack:

• Langflow is under active attack. CVE-2025-3248 (unauthenticated RCE) was added to the CISA KEV catalog and exploited to deploy the Flodrix botnet; CVE-2026-5027 (path-traversal → RCE) is being exploited across ~7,000 exposed instances.
• LangGraph checkpointer CVEs — SQL injection and unsafe deserialization in the SQLite/base checkpointers can chain to RCE via forged checkpoint rows.
• Supply-chain attacks on the CI actions and packages the ecosystem depends on (the tj-actions and LiteLLM incidents) mean a pinned-to-a-tag action or an unpinned dependency is a real risk surface.

These are exactly the things a framework-specific linter can catch before you ship. langdoctor ships the CVE data with the package and runs fully offline.

What it checks

Severities for CVE checks are derived from the CVSS score; 🔴KEV marks Known-Exploited vulnerabilities, which are always surfaced first.

CI integration

GitHub Action

# .github/workflows/langdoctor.yml
name: langdoctor
on: [push, pull_request]
permissions:
  contents: read
  security-events: write   # to upload SARIF
jobs:
  scan:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - uses: elaz48/langdoctor@v1
        with:
          fail-on: high
      - uses: github/codeql-action/upload-sarif@v3
        if: always()
        with:
          sarif_file: langdoctor.sarif

Findings show up in your repo's Security → Code scanning tab.

pre-commit

# .pre-commit-config.yaml
repos:
  - repo: https://github.com/elaz48/langdoctor
    rev: v0.1.0
    hooks:
      - id: langdoctor

GitLab CI

langdoctor:
  image: python:3.12-slim
  script:
    - pip install langdoctor
    - langdoctor --format sarif > gl-langdoctor.sarif
  artifacts:
    reports:
      sast: gl-langdoctor.sarif

Configuration

Configure defaults in pyproject.toml; command-line flags take precedence.

[tool.langdoctor]
fail-on = "high"            # critical | high | medium | low | never
ignore  = ["LD203", "LD302"]
exclude = ["examples", "vendor"]

Suppress a single finding inline — by LD id, CVE, or any alias:

API_KEY = get_key()  # langdoctor: ignore=LD401

langgraph==1.0.5  # langdoctor: ignore=LD102

Suppressed findings are always reported as a suppressed: N count so nothing disappears silently.

Output formats

--format console (default) · json · sarif · markdown

• sarif — GitHub code scanning; security-severity carries the real CVSS.
• markdown — paste straight into a PR or issue.
• json — machine-readable, with a per-severity summary.

Exit codes: 0 clean (or below --fail-on), 1 findings at/above threshold, 2 scan error. Heuristic checks never affect the exit code unless --strict.

What langdoctor is NOT

• ❌ Not a runtime guard. It scans code and dependencies; it does not sit in your request path.
• ❌ Not an LLM firewall. It does no prompt-injection or content analysis.
• ❌ Not a replacement for Semgrep/Snyk. It's the framework-specific layer those miss — LangGraph/LangChain/Langflow footguns, by ID, with fixes.

It is deterministic (no AI calls), offline (CVE data ships in the package), and zero-config.

Development

pip install -e ".[dev]"
pytest
ruff check .

Requires Python 3.10+. See CLAUDE.md for design principles, the check-ID conventions, and the "advisories are data" model.

License

MIT © 2026 elaz48

_{Topics: langgraph · langchain · langflow · security · sast · ai-agents · llm-security · devsecops · cli}