DualSentinel: A lightweight detection framework for targeted attacks on black-box LLMs based on the dual entropy lull pattern

Researchers propose DualSentinel, a lightweight defense framework for black-box LLMs, to detect during generation whether targeted attacks such as backdoors and prompt injection have been activated. The study observes that when a targeted attack successfully hijacks the generation process, the model exhibits an "Entropy Lull" in which token probability entropy is abnormally low and stable, manifesting as generation along a fixed path rather than creative selection. DualSentinel first flags this pattern at inference time through a monitoring method that is aware of magnitude and trend, then triggers a second verification based on task flipping; an attack is confirmed only if the entropy lull pattern persists in both the original task and the flipped task. The evaluation claims the framework has high detection accuracy, near-zero false positives and negligible additional overhead, and the source code has been made public.