Scalable Hierarchical Attention Transformers for Multi-Turn Jailbreak Detection in Long Conversations

The paper introduces a scalable hierarchical attention transformer model designed for detecting multi-turn jailbreaks in conversations, framing the problem as conversation-level classification rather than turn-level moderation. This model efficiently encodes individual dialogue turns into compact representations and employs a lightweight conversation module that enhances cross-turn reasoning without relying on long-context concatenation. Achieving an F1 score of 0.9394 on a benchmark of 14,038 conversations, it outperforms the Claude Opus 4.7 baseline while reducing the false-positive rate, making it a significant advancement for practitioners focused on robust moderation in conversational AI.