New detector spots sneaky jailbreak attempts hiding in normal-sounding text

Original: Detecting Fluent Optimization-Based Adversarial Prompts via Sequential Entropy Changes

Source: arxiv.org ↗