How Tutela inspects AI requests
Documentation for Tutela v3.9.0.
Tutela applies controls to requests sent through its supported connections. Understanding those connections helps you distinguish a configured policy, a recorded decision, and an action that actually completed.
Prerequisites
- A list of the applications and clients your organization intends to connect.
- Your deployment operator or administrator when reviewing network access and storage.
Follow a request
- A person, application or agent sends a request through a configured Tutela connection: a supported browser extension, AI/API gateway or MCP gateway.
- Tutela identifies the caller using the connection's authenticated identity and evaluates the applicable controls. Depending on the connection and policy, the result can allow, warn, redact, block, or require approval.
- The permitted request or tool call continues through the supported connection. Its decision and recorded execution result are separate: permission to proceed does not prove that the provider or tool completed the action.
- Activity and Evidence show recorded decisions and supporting identifiers. AI Spend uses reported usage and pricing information. If a record or source is missing, the outcome is unknown; an external action may still have occurred.
Know what each connection covers
| Connection | Intended use | Check before relying on it |
|---|---|---|
| Browser extension | Supported interactions on supported AI sites | The extension is connected, the site's interception is working, and the test request produces evidence. |
| AI gateway or SDK/API | Requests from configured applications | The client points to Tutela rather than directly to the provider and uses the intended credential. |
| MCP gateway | Calls to registered tools through Tutela | The client uses the registered gateway connection and the expected tool policy applies. |
Applications discovered from telemetry appear in inventory. That observation alone does not route their traffic through Tutela. A direct call that bypasses the configured connection cannot be assumed to have received its controls.
Understand failures and data handling
When inspection is unavailable, the configured behavior can block traffic or allow it with reduced inspection. Review the affected connection's setting with the administrator; do not infer protection from a successful response. Use Inspection unavailable or blocked when this affects work.
Activity and audit views provide decision metadata, identifiers and source state rather than raw customer prompts and responses. Keep provider credentials in the approved configuration and secret storage. Review archive coverage and exclusions separately when planning recovery.
Validate
- Every intended client has an identified connection and owner.
- A real request through each connection has been compared with its recorded evidence.
- Observed inventory, policy decisions, execution results and usage totals are understood as different measurements.
Next steps
Classification in the Laya stable release
Laya supplies category scores for prompt injection, tool injection, PII, PHI, PCI, secrets, financial information and source code. Secrets appear as Credentials, and tool injection is a distinct injection subtype. Customer policies determine the action; installed capabilities do not mean every detection is blocked.
Precise detectors locate content for redaction. A model category score alone does not locate sensitive text or prove successful redaction. When a redaction decision cannot safely locate the content, Tutela withholds it and records the limitation. The retired System 1 runtime is not part of this release.