Approving an AI Agent Isn’t Controlling It: Contain the One That Goes Rogue
From publicly posted calendar Checked yesterday
About this event
Your AI agent passed review with an identity, policy, and valid token. Then it does something unexpected: reads an API key to call a provider directly, probes neighboring databases, or uploads files to unapproved endpoints. A valid credential does not guarantee good behavior, and flat networks leave clusters vulnerable. This session explores five rogue agent scenarios, from gateway bypasses to unapproved deployments. Using eBPF kernel-level visibility, we track actual agent behavior, including direct, unrouted calls. We then demonstrate containment: quarantining misbehaving agents without affecting neighboring workloads and sandboxing new agents with default-deny policies. You’ll leave with a practical plan to contain rogue agents that pass basic checks yet exhibit compromised behavior. Platform engineers and leads who will be paged when an agent misbehaves, security engineers responsible for incident response on Kubernetes, and AI engineers who want to know what happens to their agent when it steps outside beyond its boundaries Join us to discover practical containment strategies that will help you safely isolate workloads and protect your Kubernetes infrastructure as your team scales its AI agent deployments.