Learning-to-Detect

Detects unknown jailbreak attacks in large vision-language models using hidden state analysis and autoencoders, with training and evaluation pipelines for robust safety monitoring.