There's a tension at the center of security AI. To be useful, the model needs to see your environment: the code, the cloud config, the identities, the findings. For many teams, that's precisely the data they can't send to a third party's inference endpoint. Regulated industries, government, defense, and anyone with a serious data-residency obligation hit this wall immediately.
The usual resolution is to water down the product. Redact the inputs, send less context, accept worse answers. We took the other path: SecNode runs inside your boundary, the model gets full context, and the data never leaves.
Sovereignty is an architecture decision
It's tempting to treat this as a compliance checkbox, a data-processing addendum plus a region setting. It isn't. Where inference runs, where evidence is stored, and what egress is permitted are architectural choices. They have to be made before the first request, not bolted on after.
Three deployment profiles
- Your cloud: SecNode runs in your own VPC or subscription. Inference, data, and evidence stay in infrastructure you control and can audit.
- On-premises: for teams that keep security tooling off the public cloud entirely, the same system runs in your data center.
- Air-gapped: no external connectivity at all. Models run on a local inference cluster; egress is zero by policy, not by promise.
The default should be that nothing leaves. Anything else is a decision you make deliberately, not one made for you.
What stays inside
Model inference runs locally, with no calls to an external provider mid-engagement. The Memory Layer, with its complete picture of your environment, lives in your infrastructure. Findings, reproductions, and evidence are stored where you keep the rest of your security data. Every agent action is logged, attributable, and exportable to your SIEM, so the audit trail is your own.
Why full context beats redaction
A model reasoning over redacted inputs is guessing around the holes you cut. Reachability analysis in particular falls apart when the identity graph is partial. Keeping everything inside the boundary means SecNode AI reasons over the real environment, and that's the only way its conclusions are specific enough to trust.