A preprint tests GIFT, a GPU system designed to isolate user data in LLM serving, and reports overhead in selected model and workload benchmarks.