My colleague, Martin Kraemer, writes an awesome article summarizing how to securely implement AI agents, Best Practices for Implementing AI Agents. It's a GREAT summary article.

https://blog.knowbe4.com/best-practices-for-implementing-ai-agents

The only thing I would add is that all of his suggestions should begin and include an extensive threat model. It's implied. But if you aren't threat modeling your AI implementation, you should. Most aren't doing it. If you aren't doing it, stop what you are doing and create and implement a threat model...on all AI. And this really even applies if you are just a user of AI. Your AI can be used against you. Threat model. Threat model. Threat model.