What access should an AI agent receive?
Give an agent the minimum records and actions needed for its assigned task. Define permitted tools, review points, logs, and a stop mechanism. Treat instructions found in documents or web pages as untrusted input.
Scroll sideways to read all columns.
| Permission | Define before connecting it | Check before release |
|---|---|---|
| Read records | Which sources and user permissions apply. | Try a record the requesting user may not access. |
| Change records | Allowed fields, validation, approval, and recovery. | Test a rejected change and a repeated request. |
| Send or publish | Approved destinations, content review, and who authorizes the action. | Confirm the workflow stops when approval is absent. |
| Use external providers | Data sent, retention requirements, credentials, and cost limits. | Inspect the data path and verify credentials stay out of outputs. |
Map information and actions separately.
List the sources the agent needs and the systems it may change, with an owner for each. Distinguish drafts from committed updates. Identify external messages and changes to other people’s access so you can set permissions, approvals, and review records.
Use scoped identities and tool interfaces.
Give the workflow a defined identity and only the tools, records, and operations its task requires. Check authorization, inputs, target records, and allowed changes when a tool runs. Keep credentials outside model-visible content and give staff a way to revoke access.
Treat retrieved content as evidence.
Documents, websites, and messages may contain instructions that conflict with the task. Treat them as source material and keep authorized instructions separate. Test that tool controls reject unauthorized actions and log enough context for review.
Design review, records and recovery.
Agree which actions need human approval and show the reviewer the proposed change, sources, and consequences. Record results and failures using appropriate access and retention, without unnecessary copies of sensitive content. Plan how staff stop a run, check partial updates, and resume or reverse supported changes.
Check the controls in the intended environment.
Test permissions, blocked actions, unavailable services, and escalation in the intended deployment setup. Use confirmed security and information requirements when selecting providers, storage, and monitoring. Active K Digital can design and verify the integration and application controls, with any specialist assessment agreed in scope.
Your starting checklist.
- Inventory readable information and permitted state changes.
- Scope identities, credentials and tool authorization.
- Test conflicting content and unauthorized action requests.
- Define approval evidence, logs and stop controls.
- Exercise partial failure, recovery and access revocation.