Organizations deploying LLM-based applications in Kubernetes face an unmonitored attack surface at the inference layer — prompt injection and agent manipulation — that bypasses all traditional network, file, and system call-based security controls. CrowdStrike’s Falcon AIDR capability extends the Falcon Container Sensor to instrument OpenAI-compatible API calls at runtime, addressing this gap without proxy insertion. This is a structural security architecture advisory, not a patchable vulnerability.