Przejdź do treści

Connector setup — overview

Audience. A compliance officer or IT lead at the customer deciding which connectors to install and in what order.

Time required. 10–20 minutes per connector for the hands-on admin who has the right console access (Azure subscription owner, AWS account admin, GitHub org owner, etc.). All pages are designed so the admin can complete the install without us on a call.

What a connector does. A connector is a one-time grant of read-only access from your environment to Secruna. After install we periodically scan the named services and surface every AI system we find as an inventory entry, classified against the regulatory frameworks you have enabled for your tenant.

What a connector is not. A connector is not a sensor, an agent, or a man-in-the-middle proxy. We do not see the prompts your employees send to LLMs, the data you train on, or the responses your customers receive. We only see the management-plane metadata and audit-log events that your cloud provider exposes through standard read-only APIs.

Decision tree — where to start

Pick the connectors that match the platforms you use. Order is suggestive — each connector is independent, so you can install in any sequence.

Do you run workloads on Azure?
   yes  →  connectors/azure.md           (active inventory + passive shadow detection)

Do you run workloads on AWS?
   yes  →  connectors/aws.md             (active inventory + passive shadow detection)

Do you run workloads on GCP?
   yes  →  connectors/gcp.md             (active inventory + passive shadow detection)

Do you have a GitHub organisation
where engineers push code or use Copilot?
   yes  →  connectors/github.md          (Copilot seats, AI deps, Actions, code search)

Do you use Power Platform / Power Apps /
Power Automate / AI Builder / Copilot Studio?
   yes  →  connectors/m365.md            (Power Platform admin scope)

Do you have an OpenAI organisation
(ChatGPT Enterprise, API platform)?
   yes  →  connectors/openai.md          (admin API key)

Do you have an Anthropic organisation
(Claude API platform)?
   yes  →  connectors/anthropic.md       (admin API key)

Do you have a Databricks workspace
that hosts MLflow models or serving endpoints?
   yes  →  connectors/databricks.md      (OAuth M2M Service Principal)

Active vs passive — what the difference is

Most cloud providers ship two flavours of connector:

Flavour What it does Coverage
Active Lists AI resources currently provisioned in the account (e.g., a Bedrock custom model, a Vertex AI endpoint, an Azure ML workspace). Tells you what is there now.
Passive Reads the cloud audit log for AI service calls in the past 90 days — even if those services were called from a service the active connector cannot see (e.g., a Lambda calling Bedrock without a registered model). Catches shadow AI.

When you connect Azure, AWS, or GCP, we automatically twin the connection with its passive collector so you do not have two separate flows to install. The active connector is the primary inventory source; the passive collector adds the shadow-AI coverage.

The passive collector requires one additional permission grant (documented per connector). If you do not grant it, the active connector still works — you simply do not get the shadow-AI signal.

What we do not touch

Across every connector, we never read:

  • Prompt or completion content. We see that a model was called, not what was passed to it.
  • Training data. We do not enumerate or fetch dataset contents.
  • Customer secrets beyond the credentials you explicitly granted for the connector itself. We do not list Key Vault contents, Secrets Manager entries, etc.
  • Source code outside the AI-relevant files for the GitHub connector. We use code-search to find lines that import a known AI library or call a known AI URL; we do not snapshot repositories, do not run dependency-confusion-style scans, and do not read private files outside .github/workflows/.
  • Personal data of end-users. Anything we render in the UI is metadata about an AI system (its name, the model behind it, the audit trail of who created it). End-user content of the AI system is invisible to us.

The page for each connector lists the exact API methods we call so your security team can audit the surface independently.

How to revoke a connector

The exact steps live on each connector page (look for the How to revoke access section). In summary:

  1. Revoke the grant in your console (delete the IAM Role, uninstall the GitHub App, revoke the OAuth consent, rotate the API key, etc.). This is the source of truth — once your console no longer trusts us, we cannot read anything new even if our side has not yet caught up.
  2. Disconnect the connector in Secruna (Connections → menu → Disconnect). This marks the connection inactive in our DB and stops any further discovery runs.
  3. After 90 days the artefacts collected through that connector are automatically purged. See ../legal/data-retention.md.

Cross-references