[SizzleUnrlsd]

Summary

Detectors asked for by the manual security audit of the twelve repositories under tests-project/, each traced to a line the audit cites. Three of its seven recommendations were already delivered (the call-graph crash, match, FileStorage.save); this covers most of what static analysis can do among the rest.

• Tests first (test: define the audit-driven detectors), implementation follows.
• Advisories: Gunicorn CVE-2024-1135, OpenCV CVE-2023-4863 (cv2.imread, cv2.imdecode), Werkzeug CVE-2024-34069 and CVE-2023-46136, Pillow CVE-2026-25990 (PIL.Image.open) join the offline sample with their affected APIs, so the three repositories pinning those versions report vulnerable-dependency, and the image decoders reachable-vulnerability.
• syntax/flask_debug: app.run(debug=True) is debug-enabled (high), at module level too, where two repositories start their server; resolved through the symbol table, so a run(debug=True) on another object is ignored.
• syntax/missing_timeout: Requests and httpx calls without a timeout, on the module, on sessions and on clients, are missing-timeout (low), resolved through the call graph symbols.
• The output of a subprocess is an untrusted process-output source: subprocess.run(["sh"], input=curl_result.stdout) in Chatbot's setup.py is a command injection.
• Secrets: credentials embedded in a URL literal (user:pass@ or password=/token= query parameters) are a hardcoded-secret of provider url, parameter defaults are bound to their parameter name, and the fallback of os.getenv("X_TOKEN", "literal") is bound to the variable name, which makes InfluxDB's default admin token a hardcoded-credential.
• Two Django and HTTP client fixtures gain a timeout so they keep testing what they tested.

Left from the audit for a next pass: secrets in YAML, TOML and .env files, a coverage metric per file and function, plaintext password storage, and private data served from static.