Alexander Darby: Securing AI Agents in Production

Securing AI Agents in Production Alexander Darby As AI agents become increasingly ubiquitous, the pressure to ship them quickly and keep up with the latest advancements can lead to security taking a backseat. However, this doesn't have to be the case. In this talk, we'll explore how to apply existing secure software patterns to AI agents, even in the face of rapid innovation and changing frameworks. I'll start by building an AI agent from scratch in pure Python, which will help us understand the underlying mechanics and identify potential security risks. From there, we'll discuss how to package and ship AI agents securely, using frameworks like Google ADK as an example. I'll delve into the key security challenges of AI agents, including dynamic input/output, observability, role-based access control (RBAC), and determinism vs. dynamism tradeoffs. I'll examine how to address these challenges using established software security principles, such as least privilege and validation. Along the way, I'll highlight the importance of tracking model vulnerabilities and treating them as part of the software supply chain. Ultimately, this talk will show that securing AI agents in production is not a new problem, but rather an extension of existing software security challenges. By applying established principles and patterns, we can build secure AI agents that meet the needs of modern applications. I'll conclude by emphasizing the importance of a multidisciplinary approach to AI agent development, involving data engineering, software engineering, and security architecture expertise. Talk room 2 (200 people) Sun 4:50 pm - 5:20 pm