This talk details challenges in incident response for AI systems, including insufficient logging, visibility, and accountability, as well as the risks of data exposure and prompt injection. We examine a case of RAG-enabled LLM and propose triaging strategies and improved IR practices for mitigation.
