version: "1.0.0" name: code-review-python description: > Security review of Python code — dangerous sinks and framework-specific pitfalls (Django/Flask/ FastAPI). Load when reviewing a Python codebase/PR, on .py source in scope, or "review this Python". Signals: requirements.txt/pyproject, Django/Flask/FastAPI, ORMs, pickle/yaml, subprocess. domain: code-review type: reference stability: learning modes: [bugbounty, defense, pentest] severity: info cwe: [CWE-94, CWE-89, CWE-78, CWE-502] tools: [semgrep, bandit, ripgrep] schema_version: 1

Python security code review

When it applies

You're reading Python source (a repo, a PR, a service). This is the language-specific companion to code-review-methodology — the exact sinks and framework gotchas to grep and trace.

Why it works

Most severe Python bugs come from a known set of sinks plus framework misuse. Grep the sinks, trace each argument to a user source, and check the framework's safe-vs-unsafe API was used.

Sinks & patterns (grep, then trace to user input)

• Command exec: os.system, subprocess.*(..., shell=True), os.popen — shell=True + user input = injection.
• Code eval: eval, exec, pickle.loads, yaml.load (without SafeLoader), marshal — deserialization/eval RCE.
• SQL: raw/f-string queries, .raw(), .extra(), cursor.execute("... %s" % x) — use params, not formatting.
• SSRF: requests.get/urllib/httpx on a user URL (→ web-ssrf).
• Path/upload: open/send_file/os.path.join with user paths (traversal); zip extraction (zip-slip).
• Template (SSTI): render_template_string, Jinja from user input (→ web-ssti).
• Secrets: hardcoded keys/passwords; DEBUG=True in prod.

Framework specifics

• Django: mark_safe/|safe (XSS), .raw()/.extra() (SQLi), DEBUG=True (info leak),

SECRET_KEY exposure (session forgery), missing @login_required/object-level checks (IDOR), ALLOWED_HOSTS='*', pickle session serializer.

• Flask: render_template_string (SSTI), debug=True (Werkzeug console RCE), weak SECRET_KEY

(session tampering — flask-unsign), send_file traversal.

• FastAPI: missing dependency-injected auth on routes, over-broad CORS, Pydantic not enforcing

server-side authz (BOLA at the data layer).

Method

rg -n "shell=True|eval\(|exec\(|pickle.loads|yaml.load\(|render_template_string|\.raw\(|\.extra\("; run bandit -r . and semgrep --config auto; triage by exploitable source→sink.

Gotchas

• yaml.safe_load and parameterized ORM queries are the safe variants — confirm which is used.
• Bandit is noisy; rank by user-controllable input reaching the sink.

References

Bandit; Semgrep Python rules; OWASP Django/Flask cheat sheets.