SOLUTIONS · 12
In your data centre, or in the Cloud.
The deployment model does not change how the product behaves: same engine, same policy, same evidence. What changes is where the components run and who operates them. On-prem for banks and organisations that must host in-house; Cloud for a fast start.
THE PROBLEM
A security layer is also a dependency
If a component that sits in front of every LLM request goes down, traffic either stops or flows uninspected. Which of the two happens should be a deliberate choice per application, and it should be recorded every time.
For organisations that must host in-house, every outbound connection is a separate approval. The security product itself should not call out to external services, font servers or a licence server.
HOW IT WORKS
Step by step
On-prem installation
Installed on Kubernetes or OpenShift with a Helm chart. An air-gapped profile is available for environments without internet access; images and models are pulled from an internal registry.
Separation at build time
On-prem binaries do not contain signup, billing or operator panel code. These surfaces are separated by build tags and verified by a build check.
Identity
Keycloak connects to your existing identity provider over OIDC. Console roles come from the eight-role permission model.
Egress control
Outbound connections are restricted in two layers: an egress allowlist in the application and NetworkPolicy rules open only to provider addresses.
Failure mode
Each application chooses to fail open or closed. Closed means requests are refused when a dependency is down; open means they pass and are recorded with the FAIL_OPEN flag. If the data plane cannot reach the control plane, it keeps running on the last signed policy.
Cloud
A dedicated organisation is created for you in the environment Wardline operates. The operator panel runs on a separate network with mandatory MFA and reads only directory information.
DECISIONS
Which decision you get, and when
| Decision | When | Example reason code |
|---|---|---|
| ALLOW | The classifier is unreachable and the application fails open: the request passes with the FAIL_OPEN and DEGRADED flags. | POLICY.FAIL_OPEN |
| DENY | Same situation, application fails closed: the request is refused. | POLICY.FAIL_CLOSED |
| DENY | The policy bundle has gone stale and policy says to refuse in that case. | POLICY.BUNDLE_STALE |
EXAMPLE
The classifier during maintenance
Classifier nodes in the data centre go into maintenance. The mobile assistant is configured to fail open: its requests keep being evaluated by L0 rules and policy and are marked with FAIL_OPEN and DEGRADED. The credit analysis agent fails closed, so its requests are refused during the maintenance window. Afterwards the flagged decisions are filtered and reviewed in the console.
| Decision | ALLOW |
|---|---|
| Application | mobile-assistant · fail open |
| Dependency | classifier · deadline exceeded |
| Layers that ran | L0 · policy |
| Policy | last signed bundle |
| Reason codes |
|
| Flags |
|
| Risk score | 0.00 |
CONFIGURATION
What administrators control
Installation profile
Resource sizes, high availability and the air-gapped profile are chosen through Helm values.
Egress allowlist
Which provider addresses may be reached; nothing outside the list gets a connection.
Failure mode
Open or closed, per application and per endpoint.
Identity provider
Your existing directory and MFA policy.
RELATED
Capabilities that work alongside
See it on your own traffic.
In a demo session we walk through this flow with examples your security team picks, in the console’s evidence viewer.
Request a demo