GPU Farm Hijacking

GPU compute is the scarcest and most expensive resource in the AI ecosystem — a single H100 cluster can represent millions in capital expenditure, and cloud GPU instances cost $2–$30/hour. GPU farm hijacking exploits the gap between the value of these resources and the security of the interfaces that control them. The attack surface includes: exposed inference endpoints (Ollama on port 11434, vLLM on 8000) running without authentication, Kubernetes GPU operators with default credentials, NVIDIA container runtime vulnerabilities enabling container escape to host GPU access, and stolen cloud credentials (API keys, service accounts) granting access to GPU-backed instances.

• Monitor for unexpected GPU utilization patterns (sustained high utilization outside training schedules, utilization on instances not running ML workloads)
• Alert on new GPU instance provisioning from unusual geolocations or at unusual times
• Network monitoring for connections to known cryptomining pools or unauthorized NCCL/ZeroMQ traffic
• Kubernetes audit logs for GPU resource requests from unexpected service accounts

GPU farm hijacking provides compute resources that enable T14-AT-003 (Cost Inflation) when the attacker runs workloads on the victim's account, T14-AT-009 (Resource Starvation) when hijacked GPUs are no longer available for legitimate use, and T14-AT-013 (Economic Espionage) when GPU memory access reveals model weights or training data.