AI Agent Security Checklist

AI agent security checklist

Use this checklist before exposing an AI agent to customers, production data or privileged tools. The objective is to reduce what an attacker can influence and limit the damage if the model behaves unexpectedly.

1. Minimize permissions

• Give every tool only the permissions required for its task.
• Prefer read-only access where write access is unnecessary.
• Separate high-risk actions from routine operations.
• Avoid sharing master API credentials with the model layer.

2. Treat model output as untrusted

• Validate tool parameters outside the model.
• Use allowlists for destinations, commands and resources where practical.
• Require confirmation for destructive or financial actions.

3. Protect sensitive information

• Do not place secrets in prompts unless absolutely necessary.
• Redact credentials from logs and reports.
• Limit access to private conversation or customer data.
• Test whether the agent can reveal system prompts or hidden context.

4. Test hostile instructions

• Run direct prompt injection tests.
• Test instructions embedded in external content.
• Try multi-step and encoded attacks.
• Check whether repeated attempts change behavior.

5. Monitor production behavior

• Track security-relevant actions and score changes.
• Alert on unexpected tool usage or repeated failures.
• Review significant configuration changes.
• Retest after changing models, prompts, permissions or integrations.

Security is a lifecycle

A checklist is a starting point, not a one-time certification. AI agents change when prompts, models, retrieval sources and tools change. Security testing should therefore be repeated as part of normal development and deployment.