Cerium-Delta is a real-time neural network observability platform. It extracts and analyzes internal model states: activations, gradients, weights, and information flow. Security is critical because it handles model files and numerical data that can lead to code execution if deserialized unsafely.

DO NOT open a public issue for security vulnerabilities.

Report privately:

1. GitHub: Security > Report a vulnerability (Preferred)
2. Email: helloiamnew.main@gmail.com

Include:

• Description + impact
• Repro: model type (PyTorch / TF / JAX / ONNX), Cerium-Delta version, python version
• Sample script or model that triggers it (if possible, dummy weights)
• Whether it requires untrusted model file

In Scope (high priority):

• Arbitrary code execution via model loading (pickle, torch.load, TF SavedModel)
• Path traversal in exporters / model-state extraction
• ReDoS / OOM crash via crafted activation tensors / gradients
• Data exfiltration from observability server / dashboard
• Prototype pollution / injection if JS visualization layer exists
• Supply chain: dependencies that allow code exec during install

Out of Scope:

• Model accuracy / metric correctness (that's a bug, not security, unless it leaks data)
• DoS via training huge models (expected resource usage)
• Vulnerabilities in PyTorch / TensorFlow themselves (report upstream)

WARNING: Never load untrusted models with torch.load without weights_only=True. Cerium-Delta analyzers must treat all model files as untrusted by default.

1. Acknowledge in 48h
2. Triage + reproduce in 5 days
3. Fix on private branch
4. Release + GHSA Advisory + CVE if needed

We follow 90-day coordinated disclosure.

Good-faith research is allowed. Do not access other users' models/data, do not degrade demo servers.

• Run Cerium-Delta in isolated env when analyzing untrusted models
• Do not expose real-time observability dashboard publicly without auth
• Pin versions, enable Dependabot
• Treat activation dumps as sensitive - they can leak training data

Security fixes via GitHub Releases and Security Advisories.

Maintained by NEURAL-Y