Research
Zero-shot LLM Jailbreak Detection via Internal Discrepancy ...
Publication date unknown · Discovered aclanthology.org
Page published
Publication date unknown · First observed: 10 Nov 2026
Coverage timeline
Single-source research — one report is available.
Why it matters
ALERT addresses defenders’ need to detect emerging jailbreak patterns without relying on attack templates available during detector training.
ALERT is a proposed zero-shot LLM jailbreak detector that amplifies differences between benign and jailbreak prompts across internal layers, modules, and tokens, then applies two complementary classifiers. Experiments on three safety benchmarks report strong detection performance without jailbreak templates in training, with ALERT consistently ranking among the top two evaluated methods.