Where is my data processed?

Inside your own infrastructure. The runtime and the inference sidecar both run in your VPC or on-prem, which is what makes data residency a deployment choice rather than a vendor promise - you pick the region because you own the compute.

How long is anything retained?

Retention is set per plan on the hosted console and covers audit records, telemetry, and traces separately. Enterprise adds configurable data retention policies. A self-hosted runtime retains only what you configure it to write.

Is ZNYX a data processor or a controller?

For the hosted control plane, Zitrino acts as processor on your behalf, under the Data Processing Agreement, which sets out scope, subprocessors, security measures, international transfers, breach notification, and audit rights. For a fully self-hosted deployment there is no processing by Zitrino to cover.

Data handling

Where the AI security platform runs, and what is allowed to leave your boundary.

On every plan, detection, agentic evaluation, and judge calls run inside your environment. The runtime AI firewall and the in-VPC inference sidecar keep prompt, output, and tool-payload bodies in your boundary. The hosted control plane is a separate operating surface, and Growth is designed around operational metadata by default rather than prompt and response bodies.

Data-handling model by plan for the runtime, the in-VPC inference sidecar, and the control plane.
Plan	Runtime location	Control plane location	Hosted visibility by default	Notes
Open Source	Customer infrastructure	No control plane	Not applicable	Self-hosted runtime and in-VPC inference sidecar with local policy files. No account, no metering.
Starter	Customer infrastructure	Hosted by ZNYX (free tier)	Trace id, decision, detector summary, latency, and scope metadata	The console on its free tier. The engine still runs in your infrastructure.
Growth	Customer infrastructure	Hosted by ZNYX	Trace id, decision, detector summary, latency, and scope metadata	Hosted control plane designed around metadata-first operational visibility by default.
Enterprise	Customer infrastructure	Hosted by ZNYX or customer infrastructure	Deployment-dependent	Use the self-hosted control plane and air-gapped deployment when private operations are required.

What never leaves your environment

The runtime AI firewall and the in-VPC inference sidecar run inside your boundary, so detection, agentic evaluation, and judge model calls all process content where your data already lives. Nothing below is sent to the hosted control plane.

Prompt, output, and tool-payload bodies evaluated by the runtime AI firewall
Retrieval, agent-plan, agent-step, and memory-write evaluation stages
ML detector and LLM-judge model calls served by the in-VPC inference sidecar
Embeddings, vectors, and redacted content handled by the egress and DLP gate

What the hosted control plane sees

On Growth, the hosted control plane is metadata-first by default and is designed to retain:

Trace identifier
Allow, block, redact, or warn outcome
Detector summary and policy scope
Latency and operational status
Project or environment metadata

If you need the control plane to live entirely inside your own boundary, that is an Enterprise deployment, not a hidden default of Growth.

Full disclosure

The two times content reaches ZNYX

Evaluation runs in your infrastructure on every documented integration path. These are the only exceptions, and both are things you control.

Features you use inside the console

The policy playground evaluates the text you type into it, and it is evaluated by the hosted console rather than by your runtime, because that is where your draft policy lives. Playground content is not persisted. This applies only to text you paste into the console yourself, never to your application traffic: there is no ZNYX-operated endpoint your production app can send prompts to.

An anonymous install ping from the SDKs

The client SDKs report a first-run install event: a random install id, the SDK and language versions, OS and architecture, and a run count. No prompts, responses, tool arguments, keys, or tenant data. It is on by default, off with ZNYX_TELEMETRY=0, and it fails silently when there is no route out. The runtime itself sends nothing to reach a decision.

Legal and procurement follow-through

The privacy and DPA pages cover roles, retention, and contractual handling at a high level. Executed agreements control the specifics of data processing for commercial customers.

Compare Plans Privacy Privacy Contact

FAQ

Data handling questions

What stays in your boundary, what leaves it, and how long anything is kept.

On a self-hosted runtime with a local policy file, nothing. If you use the hosted control plane, it receives operational metadata: trace identifiers, the policy decision, a detector summary, latency, and scope. Traces are designed around those identifiers rather than full prompt and response bodies.

Where the AI security platform runs, and what is allowed to leave your boundary.

What never leaves your environment

The two times content reaches ZNYX

Features you use inside the console

An anonymous install ping from the SDKs

Legal and procurement follow-through

Data handling questions

What data leaves my boundary?−

Where is my data processed?+

How long is anything retained?+

Is ZNYX a data processor or a controller?+

Secure every prompt, agent, and tool call, in your boundary.