Red Hat Benchmarks AI Guardrails

Red Hat tested nine AI guardrail configurations on prompt injection and content safety using NVIDIA's toolkit. Their DeBERTa-based classifier achieved 89.01% accuracy with 54.1 milliseconds latency, nearly matching Qwen3.6-35B's 89.31% accuracy but running significantly faster than the 312.5 milliseconds required by the larger model. Red Hat plans to ship both classifiers as defaults in OpenShift AI 3.6.