Skip to content
InfoResearchPeer-reviewedLLM-specific

Adversarial prompt detection via contrastive dual-encoder transformer for zero-day jailbreak defense based on AdverSentinel-LLM model

Published
Record updated
View JSON

Summary

Researchers present AdverSentinel-LLM, an adversarial prompt detector built on a Contrastive Dual-Encoder Transformer (CDET) that flags malicious prompts in real time, including zero-day jailbreaks. The method pairs a Prompt Semantic Encoder with a Threat Pattern Encoder, trained under a supervised contrastive objective, and adds a Dynamic Threat Memory Bank and an Anomaly-Aware Gating Mechanism. On AdvBench, JailbreakBench, ToxicChat and PromptSafety-Bench it reports an F1-score of 0.967, 97.3% detection accuracy and a 94.8% zero-day detection rate.