OSS + ELC EU 2026

"Life, uh, finds a way." So do AI agents. Give one a tool, and a prompt-injected README will teach it to spawn a shell. CVE-2025-59528 (CVSS 10.0) and Google's Antigravity sandbox escape proved the same thing twice in 2025: the agent didn't break the sandbox , either the sandbox was missing, or the electric fence was in the wrong place.
This talk is a tour of Jurassic Park, rebuilt on Kubernetes. The paddock walls are containers. The electric fence is kubernetes-sigs/agent-sandbox, swapping gVisor for Kata Containers under the same CRD. The control room is OpenTelemetry where invoke_agent, execute_tool, and tool.policy_check spans turn every isolation verdict into queryable telemetry.
We will deploy OpenClaw ,an open-source AI agent gateway inside a Sandbox resource on Cluster API, watch the agent attempt three different escapes, and ship policy decisions as first-class spans. Help shape OpenTelemetry semconv #3583 before the park opens.