Security playbook / ZARIS resources

AI agent security checklist

Review information boundaries, tool permissions, untrusted content and recovery before giving an assistant access to business systems.

The practical takeawayProtect actions at the system boundary. Instructions to the model complement, but do not replace, server-side authorization and validated tool contracts.

Start with the permitted actions.

Document every tool the agent can call and the information each tool needs. Separate reading, drafting and execution. A customer-facing assistant should not inherit an administrator’s broad permissions because that account was convenient during a prototype.

Use scoped credentials and enforce authorization in the system performing the action. Decide which operations require identity checks, human review or a second control before execution.

  • Named tool and allowed operation.
  • Validated input schema and output contract.
  • Minimal credential scope.
  • Explicit deny and escalation behavior.

Treat external content as information to evaluate.

Messages, retrieved documents and other external material can contain instructions that conflict with the intended workflow. The system needs to distinguish that content from the authority granted to the agent.

OWASP’s prompt-injection guidance is a useful reference for this class of risk. No single prompt or filtering rule establishes complete protection. Combine information boundaries, tool controls and realistic adversarial tests for the actual application.

Review what can leave the system.

Trace the data sent to models, APIs, logs and notification channels. Avoid including credentials or unnecessary personal information in prompts and diagnostics. Check that staff-facing sources cannot be exposed to unauthorized users.

Agree provider and retention settings against the actual deployment. Test requests that attempt to reveal internal information, use a different user’s record or obtain an action outside the assigned role.

  • Test and production information separation.
  • Access checks on retrieval and tools.
  • Log access and retention decisions.
  • Operator controls to revoke access or pause the workflow.

Test recovery and ongoing change.

Include tool failure, duplicate events, malformed input and uncertain intent in the acceptance set. An agent must not report a completed business action when the connected system has not confirmed it.

Name the owner for monitoring, incident response and permission changes. Re-test after material changes to tools, knowledge, providers or authority. This checklist supports a review; it is not a certification or a guarantee that a deployment is secure.

General workflow guidance. Product availability, provider requirements and implementation scope are confirmed for each project.

Further reading

Source documentation.