Security

Designed so the AI cannot hurt you.

The rules that decide what is blocked are deterministic and human-approved. The models only help you find and write them.

Only approved rules can block

• Three tiers raise alerts: statistical anomaly detection, a layer-7 classifier, and the Claude advisor. None of them can enforce.
• Only the policy engine blocks, and only through a rule a person approved. This boundary is enforced in the code base by an import contract, not by convention.
• An AI-drafted rule is schema-checked, forced to start alert-only, and replayed against recent traffic before a person ever sees it.

Signed and verifiable

• Rule bundles are signed with Ed25519. Nodes verify every bundle, refuse replays, and keep the last known good one.
• Every change, approval and sign-in goes into a hash-chained audit trail that you can verify from the console.
• The audit trail records opaque user ids, not email addresses, so a person can be erased without breaking the chain.

What it sees

• NeuraWall analyses flow metadata, TLS fingerprints and DNS behaviour. It does not intercept or decrypt TLS, and payloads are not stored.
• Claude receives flow metadata as delimited, untrusted data under a call budget. Text inside it is never treated as instructions.
• Without a Claude key, an offline advisor runs on your own host and nothing leaves it.

Access and operations

• Roles from viewer to admin, with optional four-eyes approval. Sign-in through NEXORA creates viewers only; roles are never taken from the identity provider.
• Outbound webhooks are signed and SSRF-guarded: every address is checked, the connection is pinned and redirects are not followed.
• Plan changes never change what nodes enforce, and nodes keep enforcing when the control plane is unreachable.

Report a vulnerability

Follow the disclosure process in NEXORA's security.txt. See also the NEXORA security architecture.