Adaptive Multi-Tier Inspection for Prompt-Injection Detection in LLM Gateways: Latency, Recall, and False-Positive Trade-Offs

Abstract

This study presents TrustGuard, an adaptive multi-tier security gateway for prompt-injection detection in Large Language Model (LLM) applications. The system combines a low-latency heuristic detector with an in-process dense-vector semantic detector and dynamically routes requests based on calibrated risk scores. Using a 220-sample benchmark covering 11 prompt-injection attack categories and six benign domains, the study evaluates detection recall, latency, false-positive rate, router starvation, prototype generalization, and adversarial robustness under a strict zero-leakage evaluation protocol. Results show a clear trade-off between inspection latency and detection recall. The study identifies Router Starvation as a key failure mode in adaptive cascades and evaluates feature-engineered routing as a mitigation. It also reports negative results involving prototype expansion and semantic dilution under benign contextual padding. The accompanying work provides reproducible experimental configurations, benchmark methodology, and implementation details for evaluating adaptive prompt-injection detection gateways.