When AI Goes Awry - Responding to AI Incidents

Learn how to effectively respond to AI system incidents in this 48-minute conference talk from BSidesSF 2025. Explore the unique challenges that arise when conducting incident response for AI systems, including issues with insufficient logging, limited visibility, and unclear accountability structures. Discover the specific risks associated with AI deployments such as data exposure vulnerabilities and prompt injection attacks. Examine a detailed case study involving a RAG-enabled Large Language Model to understand real-world AI incident scenarios. Gain practical knowledge about triaging strategies specifically designed for AI incidents and learn improved incident response practices that can help mitigate AI-related security issues. Understand how traditional incident response methodologies need to be adapted to address the complexities and novel attack vectors present in AI systems.