Base Models Can Detect Hidden Intentions that Prompted Classifiers Miss

Scoring a model's responses under a base model conditioned on a description of a hidden intention catches hidden prompts that prompted classifiers miss.