The Attacker’s Lens: Threat Modeling for AI and Agentic Systems

Abstract

Most teams building with AI are using security playbooks written before AI existed. That’s a problem. Modern AI systems don’t just answer questions — they browse the web, run code, read files, and take actions on your behalf. That opens up a whole new category of attacks that traditional security thinking doesn’t cover.

This talk breaks down the real threat landscape for AI-powered systems based on hands-on experience building and auditing them. The session covers where attackers are most likely to get in, what most builders are missing, and how to start thinking about AI security without discarding everything you already know.

Key Takeaways

1. A simple, layered way to map out where your AI system is vulnerable — covering inputs, memory, tools, and outputs — that you can start using right away

2. The specific attacks most teams aren’t taking seriously yet: malicious content sneaking in through retrieved data, manipulated conversation context, and chained tool abuse

3. How to add AI security review to your existing development process without building a whole new program from scratch

4. The mindset shifts that matter most when thinking like an attacker against an AI system