KubeCon + CloudNativeCon India 2026

LLMs can write code - and sometimes running that code is the most direct way to deliver product value. The moment you do, you’ve effectively introduced a remote-code-execution surface: the code is untrusted by default, but the system still has to execute it to stay useful.

In this talk, we’ll share what it took to build and operate production sandboxes for LLM-generated code at scale. We’ll cover the isolation model (containers, least-privilege defaults, syscall/filesystem restrictions), the operational reality (startup latency, resource limits, cold starts, observability), and the guardrails that matters when code or users try to misbehave. We’ll also dig into data protection: locking down egress, blocking exfiltration paths, and keeping secrets out of reach.

We’ll cover what worked, what failed, and what we’d do differently - ending with a practical, vendor-agnostic mental model and checklist you can apply.