The tool never sees your prompts
Inspection happens where your data already lives. Adopting ZNYX does not add a third party to your data-flow diagram.
Self-hosted AI security
The security tool should not be the thing that moves your data.
Most AI guardrails require you to send prompts to a vendor to have them inspected. ZNYX inverts that: the engine ships to your boundary, and nothing about a decision requires an outbound call.
Judge models stay inside too
The inference sidecar runs ML classifiers and the LLM judge in your VPC, so escalation is not a back-door egress path.
Auditable by reading it
The detection runtime is open source. Your security team can review exactly what executes rather than trusting a description.
Residency is a deployment choice
Run it in the region you are required to run in. No dependency on where a vendor happens to host.
The stack
What runs where
| Component | Responsibility | Location |
|---|---|---|
| Runtime | Detection engine and policy enforcement · Docker or Kubernetes | Your VPC |
| Inference sidecar | ML classifiers and LLM judge models | Your VPC |
| Policy bundles | Versioned, signed, pinned per environment | Your VPC |
| Trace store | Decisions, detector verdicts, audit events | Your VPC or control plane |
| Control plane | Console, policy authoring, evidence exports | Hosted or self-hosted |
FAQ
Self-hosting questions
What self-hosting gives you, what it costs to run, and where the hosted console fits.
The detection engine runs on compute you control. Prompts, responses, retrieved context, and tool payloads are inspected in your VPC or on-prem, and no request has to leave your network for a policy decision to be made. That is the difference between a security control you operate and an API you depend on.
A container runtime. The runtime ships as a Docker image and runs under Docker Compose or Kubernetes, with a Helm chart on Enterprise. It is deliberately dependency-light and needs no GPU, because the deterministic layer is rules and pattern matching. ML detection is opt-in and adds an inference sidecar.
Yes. The runtime and the engine are open source, with no account, no quota, and no telemetry in the decision path. You pay only if you want the hosted control plane for policy versioning, traces, benchmarks, and audit evidence.
Yes. The runtime reaches a decision with no outbound call, models are fetched explicitly and pinned by digest rather than downloaded implicitly at runtime, and Enterprise can self-host the control plane so no ZNYX-operated component sits in the path. Set ZNYX_TELEMETRY=0 to suppress the SDKs' install ping.