← All integrationsINTEGRATION · AI PROVIDERS

Kapālins for Azure OpenAI

GPT models deployed in your own Azure subscription. Kapālins routes each call to your deployment by name, checks the prompt and the reply on the way through, and reads Azure Monitor for the calls that went straight to the resource.

See it governed liveSetup in four steps

Your Azure resource, one policy on top

Calls go to the resource endpoint you own, with its key sealed in the Kapālins vault. The same redaction, injection defence and per-team policy as every other provider apply on the way.

Deployments nobody registered

Connect with a read-only service principal and Kapālins lists every deployment on the resource, so a model someone deployed last week is on the list, not a surprise.

Direct calls still land in the timeline

The Azure Monitor connector reads Azure OpenAI diagnostics and AI Foundry audit logs from Log Analytics, so traffic that bypassed the gateway is still seen.

Two planes, one policy

Inline on every call. Azure Monitor for the rest.

INLINE · AI GATEWAY

Every governed call, checked both ways

Your application calls the gateway. Kapālins sends the request to the deployment on your resource with its api-key, streams the reply back, and inspects both directions.

  • Deployment names route as model names
  • Native streaming
  • The same virtual key as every other provider
AFTER THE FACT · AZURE MONITOR

Every call the resource logged

The connector queries your Log Analytics workspace with a read-only service principal and turns each record into the Kapālins event shape, beside the gateway’s verdicts.

  • Azure OpenAI diagnostic logs
  • AI Foundry audit logs
  • The same timeline, incidents and evidence chain
Deployments the gateway has never seen are found separately by the Azure inventory on the Microsoft side of the estate.
ONE DOOR

One endpoint, one key, every provider.

Point the application at your gateway host with a virtual key. Kapālins routes to Azure OpenAI with the key it holds, fails over when Azure OpenAI is down, and meters every call to the team that made it.

  • Your app speaks OpenAI API format
  • Spend metered to the team that made the call
  • Block the provider domains at your firewall and this becomes the only way out
Kapālins AI Gateway: one endpoint, virtual keys and the connected providers
The gateway in the pilot tenant on production: one endpoint, virtual keys, and the providers behind them. Key prefixes masked.
FOLLOW ONE CALL

A blocked Azure OpenAI call, opened.

Open any call to your deployment in Logs and Kapālins shows who called, the deployment it was headed for, what was inspected and which guardrail stopped it, before the prompt reached your resource.

  • A scrubbed preview or only a SHA-256 hash, your choice
  • The caller’s recent history beside the verdict
  • The same record feeds incidents, Microsoft Sentinel and the evidence pack
Kapālins Logs: a real Azure OpenAI call blocked, opened to its kill chain
Recorded in the pilot tenant · production
A real test call through the Kapālins gateway to an Azure OpenAI gpt-5-mini deployment, blocked in the pilot tenant on production on 18 October 2026 because it carried a Medicare number. Recorded from the console; nothing edited.
AFTER THE FACT

Azure Monitor, in the same timeline.

An app that calls the resource directly never touches the gateway, but the resource still logs it. Kapālins reads those records from Log Analytics and sets them beside the gateway’s verdicts.

  • Read-only service principal
  • Source label on every row
Kapālins Logs: Azure Monitor records beside gateway verdicts
The Kapālins timeline, where Azure Monitor records land beside gateway verdicts. Pilot tenant on production; personal addresses masked.
AGENTS

Agents calling Azure OpenAI, fingerprinted.

Every agent and service key that calls Azure OpenAI through the gateway gets a behaviour baseline: calls, active hours, models, prompt repetition, injection rate. When one drifts it is flagged with the reasons, and one click drafts a guardrail or quarantines the key.

  • The tools and MCP servers each agent actually invoked
  • Detect only by default; quarantine is one click and reversible
Kapālins Agentic Behaviour: an agent key flagged, with its score derivation and behavioural fingerprint
Console replay · illustrative data
A replay of the Kapālins console: an agent flagged for injection activity, its score explained line by line, its fingerprint against baseline, and the MCP calls it made outside its declared scope. Agent names and activity are illustrative.
MODEL ASSURANCE

Every Azure OpenAI model, approved or not.

The Probe Runner, the garak red team and Alignment Lab test each model you use, through your own key. A model is approved, conditional or blocked, and Policy Shield enforces the verdict at the gateway.

  • Injection, unsafe code, harmful behaviour and hallucination suites
  • A critical finding holds new use until someone approves it
Kapālins AI Systems: per-model approve, conditional and block verdicts by provider
Per-model verdicts in the pilot tenant, from the Probe Runner: approved, conditional or blocked before anyone relies on them.
Setup

Four steps.

  1. Collect the endpoint and keyFrom the Azure OpenAI resource: its endpoint URL and one of its keys.
  2. Add a reader principalA service principal with the Reader role, used once at connect time to list deployments.
  3. Add Azure OpenAIIntegrations › Azure OpenAI › paste the endpoint, key and principal. Deployments are listed live.
  4. Add Azure MonitorIntegrations › Azure Monitor › your Log Analytics workspace, for calls that skip the gateway.
Request a briefing

See Azure OpenAI governed in thirty minutes.

Bring the person who holds the credential. We connect it live and watch the first events land.

Request a briefing All integrations
FAQ

Questions buyers ask.

Does the gateway add latency to Azure OpenAI calls?

Policy Shield runs in the request path. Each family can be set to detect only, and the gateway reports its own added time per call in the timeline.

Do my applications need code changes?

A base URL and a key. The gateway accepts the OpenAI and Anthropic formats, so request and response handling stay as they are.

Can Azure OpenAI be governed without the gateway?

Routing is what makes inline control possible. AI that cannot be routed, such as hosted copilots, is governed from its audit logs, identity grants and the device sensor instead.