Project: The Sentinel — Network Intrusion Detection System (NIDS)

The Sentinel is an ML-powered security tool — which means we hold it to the same standard it applies to network traffic. Thank you for taking the time to report issues responsibly.

Please do NOT open a public GitHub issue for security vulnerabilities.

1. Preferred: use GitHub's Private Vulnerability Reporting — this keeps the report confidential end-to-end and lets us coordinate a fix and advisory.
2. Alternative: open a draft security advisory through the repository's Security tab.

• Affected notebook / component (01_EDA, 02_Training_GPU, 03_Inference_API, 04_Dashboard, _build/*)
• A minimal reproduction (cell, payload, or request body)
• Impact assessment — what could an attacker do?
• Any known workarounds

• The embedded FastAPI API in notebook 03 (input validation, metadata handling, WebSocket fan-out, response sanitization)
• Artifact integrity — anything that could make model.pkl / scaler.pkl / label_encoder.pkl load or behave incorrectly (e.g., bypassing the n_features_in_ parity guards)
• Notebook code paths that execute or persist attacker-controlled data (feature vectors, _source_ip metadata, replayed CSV content)
• Supply-chain issues in the pinned Colab dependencies (_build install cells)

• Google Colab platform security — report to Google VRP instead
• Denial-of-service against the demo API — the embedded server is intentionally demo-grade: optional API key (unset by default), no TLS, per-IP rate limit 120/min, WS client cap 20
• Missing authentication / TLS on the public tunnel URLs — documented limitation of the Colab edition; reports on this alone will be closed as documented behavior unless accompanied by a concrete hardening proposal
• Attacks requiring write access to your Google Drive or Colab VM — that is platform compromise, not a project vulnerability
• Findings from intentionally adversarial inputs in _build/smoke_test.py — those are the test suite's own fixtures

These are documented in README.md / docs/NIDS_PRD.md and are not vulnerabilities:

1. No user authentication / authorization — the optional X-API-Key + WS token is defence-in-depth, not access control.
2. No TLS termination on the Cloudflare quick tunnel / Gradio share URL.
3. Per-process rate limiting and WS counters (single-process demo deployment).
4. Up to 8 of 52 missing features are zero-filled by design and surfaced via missing_features.

If you discover an exploitable issue, please do not run it against public deployments you do not own. For local reproduction, use the synthetic-flow fixtures in _build/smoke_test.py rather than live traffic.

This policy follows the standard responsible-disclosure practice. For general questions (not vulnerabilities), see SUPPORT.md.