Kubernetes Secret controls cannot protect API keys already injected into pods

A security researcher demonstrated on September 13, 2026, that standard Kubernetes hardening controls fail to protect LLM API keys once they have been injected into a pod as environment variables. Despite enabling restricted Pod Security Standards, strict RBAC policies, disabled service account token mounting, and a read-only filesystem, a single file read inside the container was enough to extract the credential. The stolen key was then successfully authenticated to the LLM provider from a separate pod, confirming the breach. The researcher clarifies that Kubernetes admission and RBAC controls govern access to Secret objects in the API, but do not restrict credentials a pod has already been granted at startup. The finding is especially relevant for AI agent workloads, which are designed to act on arbitrary text instructions and therefore represent a high-value target if a credential is exposed.