Researchers evaluated five guard models for prompt classification and found that they can be overconfident when faced with adversarial attacks, leading to high-confidence errors. This mismatch between guard confidence and base model uncertainty can have significant implications for safety classifiers in AI systems.