Skip to main content

AI Safety and Production Checklist

An LLM proposes a tool call; it does not grant authority. The host application must decide whether that operation is allowed for the signed-in user and the current document.

Before executing a tool​

  • Authenticate the user and authorize access to the document independently of the prompt.
  • Maintain a per-workflow tool allowlist and reject unknown tool names.
  • Validate arguments against the published JSON schema and apply size limits.
  • Require confirmation for delete, replace-all, external sharing, or broad document changes.
  • Cap the number of model and tool steps for one request.

After execution​

  • Record actor, tenant, document identifier, tool name, outcome, latency, and correlation ID.
  • Avoid recording provider keys, authorization tokens, or complete document content.
  • Return a bounded, structured result to the model.
  • Let the user review the changes and decide whether to save.

Treat document content as untrusted​

Text inside a document can contain prompt-injection instructions. Do not allow document text to change system policy, expand the tool allowlist, reveal secrets, or bypass confirmation. Keep system instructions and authorization decisions outside model-controlled content.

Preview boundary

The Playground demonstrates the integration flow. It does not provide production authentication, authorization, user approval, quotas, or auditing for your application.