As enterprises race to deploy customer-facing and agentic AI, traditional security approaches are no longer enough. Every input is valid, small prompt variations can produce dramatically different behavior, and the possible attack surface is effectively infinite, making it impossible to benchmark or pen-test your way to safety. Agentic AI raises the stakes even further, introducing risks such as tool-calling exploits, memory poisoning, and multi-agent attacks, while responsibility for failures spans Security, Product, Legal, and Operations.
This session makes the case for a new security paradigm: AI red teaming. Rather than relying on static tests, organizations must continuously use AI to adversarially test AI before it reaches users. Drawing on Alice's experience safeguarding over a billion human-AI interactions every day, we'll examine real-world failures, expose the gaps in today's enterprise AI security stack, and demonstrate why securing AI now requires AI itself.