5 Defence Patterns Every AI Agent Needs (Before It Gets Hacked)

In Episode 1, I showed you how to break an AI agent. In this episode, I show you how to actually defend one. I walk through 5 production patterns to secure agentic AI systems — the ones that catch real attacks like tool poisoning, prompt injection, and exfiltration before they reach your tools. Built on Google's ADK, demoed against the same malicious page from Episode 1. If you're shipping agents to production and worried about what could go wrong, this episode gives you a practical playbook. ═══════════════════════════════════ 🎯 THE 5 PATTERNS 00:00 Why your agent's whole world is the attack surface 00:55 Pattern 1 — Least Privilege (limit what the agent can do) 02:40 Pattern 2 — Dual LLM Sandbox (limit what the agent can read) 04:27 Pattern 3 — Output Validation (judge before you execute) 07:06 Pattern 4 — Human in the Loop (the limits of automation) 08:38 Pattern 5 — Observability (you can't defend what you can't see) 10:44 Bringing it all together ═══════════════════════════════════ 🔗 CODE & RESOURCES → Demo repo: https://github.com/TomaIjatomi/youtube/tree/main/episodes/002-five-defence-patterns → Episode 1 (the attack): https://youtu.be/R7Y4i-yEa54 → Google ADK: https://google.github.io/adk-docs/ → OWASP Agentic Top 10: https://genai.owasp.org/resource/agentic-ai-top-10/ → Phoenix (tracing): https://phoenix.arize.com/ ═══════════════════════════════════ 📚 FURTHER READING → Anthropic's prompt injection research: https://www.anthropic.com/research → Simon Willison on Dual LLM pattern: https://simonwillison.net/2023/Apr/25/dual-llm-pattern/ → MCP security: https://modelcontextprotocol.io/ ═══════════════════════════════════ 👋 ABOUT ME I'm Toma — AI practitioner based in Dublin, MSc in AI from Dublin City University, UNESCO Women4EthicalAI Champion. I make videos about how AI systems are actually built, deployed, and secured in production. 🌐 https://tomaijatomi.com/ 💼 https://www.linkedin.com/in/toma-ijatomi/ ═══════════════════════════════════ 🔔 If this helped you, hit subscribe — Episode 3 is coming. #AgenticAI #AISecurity #LLM #PromptInjection #GoogleADK #AIEngineering