What the runtime processes
Prompts, responses, streamed tokens, retrieved context, tool payloads, and agent plans, all inside your own infrastructure. The runtime makes no outbound call to ZNYX to reach a decision.
Security & trust
The security tool should not become the exposure.
ZNYX is designed so that adopting it does not widen your data-flow diagram. Here is exactly what runs where, what we receive, and what we do not claim yet.
What the hosted console receives
Operational metadata by default: trace id, policy decision, detector summary, latency, and scope. Prompt and response bodies are not part of that payload.
Where models run
ML classifiers and the LLM judge run on an in-boundary inference sidecar, so escalation does not become an exfiltration path.
Fully private option
Enterprise can self-host the control plane as well, for an air-gappable deployment with no ZNYX-operated component in the path.
One outbound call, and how to stop it
The client SDKs send an anonymous install ping on first run: 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, ever. It is on by default and off with ZNYX_TELEMETRY=0, and it fails silently when there is no route out.
Controls
Controls a reviewer will ask about
- Authentication
- SSO / SAML, SCIM provisioning, MFA (Enterprise)
- Authorization
- 3-tier RBAC: Owner, Admin, Developer
- Data residency
- Runtime in your region by self-hosting; hosted control-plane region is a recorded preference, not enforced placement
- Trace retention
- 7 / 30 / 90 days by plan, configurable on Enterprise
- Audit retention
- 7 / 365 / 730 days by plan
- Encryption
- TLS in transit; storage encryption in the control plane
- Deployment
- Docker, Kubernetes, VPC, on-prem, air-gapped
- Source availability
- Detection runtime is open source and auditable
Honest limits
What we do not claim yet
- External certifications (e.g. SOC 2, ISO 27001) are on the roadmap and are not claimed today.
- Published latency targets are internal benchmark figures, not contractual SLOs.
- Security and performance claims on this site reflect implemented capabilities.
If a questionnaire needs something not listed here, ask us directly rather than inferring it, we would rather answer than have you assume.
Contact the security team →FAQ
Security questions
What the runtime processes, what the console receives, and how ZNYX itself is built and operated.
No. The runtime evaluates prompts, responses, streamed tokens, retrieved context, and tool payloads inside your own infrastructure, and makes no outbound call to ZNYX to reach a decision. The hosted console receives operational metadata by default - trace id, policy decision, detector summary, latency, and scope - not prompt and response bodies.
On an in-boundary inference sidecar you deploy yourself, so escalating a borderline request to a classifier or a judge does not become an exfiltration path. The runtime is rules-only until you point it at that sidecar.
Not today. ZNYX produces evidence that supports frameworks such as the EU AI Act, ISO 42001, and NIST AI RMF, and maps its controls to the OWASP LLM Top 10, but ZNYX itself does not currently hold SOC 2 or ISO 27001 attestation. Those are on the roadmap and are not claimed as shipped.
SSO and SAML with SCIM provisioning and MFA on Enterprise, and a three-tier role model - Owner, Admin, Developer - on every plan. API keys support rotation with grace periods.
Yes. The open-source runtime with a local YAML policy file needs no control plane, and Enterprise can self-host the control plane as well for an air-gappable deployment. The one outbound call to be aware of is the SDKs' anonymous install ping, which is off with ZNYX_TELEMETRY=0.