"You Can't Secure What You Can't See": Observability for AI Agents
As AI moves from hype to day-to-day reality, a new challenge has appeared: rogue AI agents built and used beyond official channels. These range from well-intentioned employee-made copilots to unvalidated third-party AI bots quietly plugged into enterprise data. This session dives into the dual nature of these agents: incredible productivity drivers and potential security nightmares. We'll unpack the key risk areas, like sensitive data leakage, lack of accountability and agents acting with unchecked privileges, showing why traditional defenses fall short. Then we'll pivot to solution, outlining Zero Trust-inspired strategies to regain control. Learn how to treat AI agents as first-class identities (unique credentials, least-privilege access) and apply continuous monitoring so that even autonomous code follows corporate security rules. We'll share best practices for logging and auditing AI-driven actions to build trust internally and with regulators. Attendees will also get a glimpse of how these concepts manifest in emerging enterprise tooling by a neutral look at Microsoft's Agent 365 as control plane and reliance on Microsoft security ecosystem for identity management, data security and monitoring/auditing capabilities.
Pouya Koushandehfar
Pouya Koushandehfar, Senior Cybersecurity Architect and Leader · PwC