AI workloads deployed in Kubernetes environments have introduced an attack surface that most enterprise security stacks cannot see: the prompt layer. Attackers can inject malicious instructions directly into large language model inputs, bypassing signature-based and behavioral detection tools entirely, with no CVE to patch and no network anomaly to catch. This represents a structural visibility gap, not a discrete software defect, and organizations scaling AI applications without prompt-layer instrumentation are operating blind to a threat class OWASP has ranked the top risk in LLM deployments.