Anyone who is working with AI, or considering it, should care about security. The traditional attack surfaces and mitigations still apply to an AI-powered system or product. New attack surfaces appear depending on the specific AI approaches used. In addition, because AI systems are typically more automated, they can do more harm, faster, when compromised.

In this talk, we show how to recognize the attack vectors AI shares with traditional software, and what those attacks look like. We then turn to attacks specific to AI: prompt injection (direct and indirect), hijacking an agent through the content it reads and the tools it calls, poisoned training data, and adversarial inputs that fool a vision model. Using a map of the AI pipeline's attack surfaces, we walk through real incidents at each stage. For each, we show what thwarts it, or at least mitigates it: validating training data, sanitizing and rephrasing inputs, and red-teaming the model. Every mitigation comes with a real cost of capability or convenience, which we describe.

No prior AI or security experience needed. You'll leave able to threat-model an AI-enabled feature or product, choose mitigations for the attacks that apply to it, and know where to look as the landscape evolves.