#608: Before You Deploy AI Agents, Understand These Attacks

Big thanks to Proton VPN for sponsoring this video. You can use my link: https://protonvpn.com/davidbombal to get 70% off your Proton VPN subscription

How do you hack an AI system? And what happens when one malicious instruction spreads between AI agents?

I’m joined by Harriet, author of Practical AI Security, to explore how AI models and agents can be manipulated, with practical Python demonstrations. You'll see an image classifier misidentify a rifle, a demonstration of prompt injection spreading across five agents, and how memory poisoning can plant instructions that affect an agent later. We discuss why securing AI takes more than blocking prompt injection, how machine learning creates different security challenges, and what you need to understand before deploying agents in your organisation.

Want to learn this yourself? Harriet explains how to get started with her collection of 30+ free Python notebooks, including examples you can explore in Google Colab. We also discuss the skills involved in AI security and how people from different backgrounds can contribute.

Topics include:

• Adversarial machine learning and image manipulation

• Prompt injection and attacks spreading between agents

• Memory poisoning and model backdoors

• AI supply chain security

• Learning AI security with Python

• AI security careers, training and fundamentals

// Link to No Starch Website for Harriet Farlow’s Book //