NeuroGrid processes household energy telemetry and federated model updates. Treat both as sensitive data.

• Insecure gRPC is restricted to loopback addresses by default.
• Non-loopback deployments must configure mutual TLS using the variables documented in .env.example.
• Certificate common names or subject alternative names must match the declared agent ID.
• Remote LLM endpoints are disabled by default. If explicitly enabled, HTTPS is mandatory.
• Federated payloads are size-limited, schema-hashed, versioned, and loaded with PyTorch's restricted weights_only loader.
• Legacy pickled scaler artifacts are rejected unless a trusted operator explicitly opts in.

Do not open a public issue containing household data, credentials, or an exploitable proof of concept. Contact the repository owner privately with reproduction steps and affected versions.

The project implements authenticated transport and robust aggregation primitives, but it does not claim cryptographic secure aggregation: District servers can inspect individual updates. Round-level Gaussian output perturbation is also not equivalent to per-example DP-SGD.

Continuous integration treats dependency-audit findings as blocking except for PYSEC-2026-139 and GHSA-rrmf-rvhw-rf47. Both affect PyTorch and currently have no published fixed release in the audit database. These exceptions are explicit and narrowly scoped; they should be removed as soon as a compatible fixed PyTorch release is available.