Segment-Level Coherence Probe Reduces False Alarms in CBRN Jailbreak Detection by Requiring Multi-Token Evidence
arXiv·medium signal
Existing streaming jailbreak probes for LLMs trigger false alarms when sensitive CBRN terms appear in benign contexts because they rely on a few high-scoring tokens. This paper introduces a streaming probing objective that requires multiple consecutive segments to agree on harmful intent before flagging, dramatically reducing false positives while maintaining detection of actual jailbreak attempts in Chemical, Biological, Radiological, and Nuclear domains.