When Agents Go Rogue: Threat Protection for Agentic AI Systems

Speakers

• Saurabh Mhaisekar

About this session

AI agents are no longer simple chatbots responding to prompts. They plan, reason, use tools, hold memory, and make autonomous decisions — often running with overprivileged identities and unchecked access to production systems. This introduces an entirely new class of security threats that traditional AI safety measures don't cover. In this demo-driven session, we'll walk through the agentic threat surface — tool misuse and exploitation, identity and privilege abuse, memory and context poisoning, cascading failures across multi-agent workflows, and rogue agent behavior that drifts silently from intended goals. You'll see how agents get tricked into calling APIs with malicious parameters, how poisoned context persists across sessions, and how one compromised agent corrupts an entire chain. Then we fix it — using Prompt Shields, Task Adherence, scoped identities, and inter-agent trust boundaries to systematically close each attack vector. Your agents are already in production. The question is — who else is using them?