AI Security Expert: Your AI Agent Is One Emoji Away From Being Hacked
Your AI agent may not need a sophisticated exploit to be hacked. Sometimes, one emoji is enough. In this episode of The AI Frontier Playbook, I sit down with Richard Diver, Director of Story Design at Microsoft, author of Guardians of AI and Unlearn Security, and a security veteran with 30 years of experience. We explore why the biggest AI security problems aren't really technology problems — they're thinking problems. Richard breaks down how threat modeling changes when AI can act autonomously, why every AI system should be evaluated through three simple questions — asset, access, and actor — and why you should think through the "what if?" before an agent ever reaches production. We get into prompt injection, indirect attacks hidden inside documents and images, AI memory as a new attack surface, and why giving one AI access to everything may be one of the biggest mistakes organizations can make. We also explore Richard's "green team" concept — why security can't remain the responsibility of a small security department, and why the people actually using AI every day need to be part of the security process. He shares six things companies should check before turning on their first AI agent, how to limit an agent's blast radius, and why the worst attacks aren't the ones that cause immediate damage — they're the ones nobody notices for years. You'll learn: Why major security failures are usually caused by how we think, not a lack of security tools Richard's Asset, Access, Actor framework for threat modeling AI systems How a single emoji or special character can become an attack vector against AI Why prompt injection can hide inside documents, emails, images, and other content Why giving one AI access to everything creates unnecessary risk Why Richard would rather have 1,000 agents doing one job each than one AI with access to everything Why "read-only" access isn't necessarily safe when an AI can read and share information How AI memory creates a new long-term attack surface What the "green team" is and why security needs to involve everyone The six things to check before turning on your first AI agent Why reducing the blast radius may be more realistic than trying to prevent every attack Why the most dangerous AI attacks are the ones that go undetected Whether you're building AI agents, deploying AI across an organization, or simply using AI tools in your everyday work, this episode gives you a practical way to think about security before something goes wrong. Resources & Links Richard on LinkedIn: linkedin.com/in/rdiver Book: Guardians of AI Book: Unlearn Security Our sponsor Visionet: visionet.com 📬 Enjoyed this episode? Connect with me on LinkedIn: linkedin.com/in/samuelboulanger Visit the website: frontierplaybook.ai Newsletter: theaifrontierplaybook.substack.com
