Ready to get started?
Deploy sovereign AI on your infrastructure - in weeks, not months.
4 minutes · 13 steps · 1 seat
One guardrail profile is bound to every department. You read what it checks, try to talk your way past it, and then follow the record of who changed what.
In this story you will
Step by step
Each seat is a person on the demo. Where the story changes hands, it says so. Under a step, the tutorial for that screen, by code, opens the written guide on the docs.
Production · platform.admin@demo.katonic.ai
Control Room → Guardrails. katonic-baseline is the default profile: every department gets it unless another is set. Three of its twelve rails are active.
Input rails run before the prompt reaches the model, output rails on the reply before it leaves, and tool rails on the arguments of a tool call.
The profile states its own limits. A rail set to Warn only logs: the content still reaches the user. Block and Redact on a streamed reply hold the whole reply until it has been checked.
On the input side, injection detection, PII detection and content safety are enabled. The output and tool rails below are listed as disabled in this profile, so you can see exactly what is not being checked.
Click Test: a bench that runs any text through this profile and reports each rail's verdict.
We typed a classic attempt to make a model drop its instructions. Change it if you like, then go on.
Click Run Test.
The verdict is Blocked, triggered by injection detection. The other enabled rails passed, and the two disabled input rails are reported as not checked. On this demo the rails are pattern checks; no guard model is pinned.
Control Room → Policies. Each policy has a status and a priority: here PII redaction and a tool allowlist are active, and one for pinning the production model is disabled.
Control Room → Approvals. A tool call that needs approval waits here for the Admin tier. One approval is needed, and a request nobody acts on is declined.
Control Room → Audit Log. Each admin action is listed with its time, the person who did it and the resource it touched: a deployment created, an environment added, a setting changed.
Export CSV hands the listed events to your own tools. What's audited says which actions are recorded and which are not.
What it proves
At the end
A default profile on every department, an injection attempt blocked before it reached a model, tool calls under policy and approval, and each admin action listed with who did it and when. The page also told you which rails are off; nothing claimed more than it does.
Take away
5 tutorials for this story
What this story sells
This story sells
The next product up
Added on the same cluster and the same licence; nothing the smaller product had is lost.
Next step
The demo is public, with no form in front of it. An engineer takes thirty minutes and brings your own question.
