Qi Cao

The paper theorizes that aligned LLMs remain jailbreakable due to 'Refusal-Escape Directions' (RED), which are continuous perturbation paths that shift model behavior from refusal to answering, and shows this vulnerability is linked to specific operator-level sources within the model architecture.

The paper introduces AIBuildAI-2, a knowledge-enhanced agent that significantly improves the automatic building of AI models by integrating an external, evolving knowledge system, achieving state-of-the-art performance on benchmark tasks.

The paper introduces SCOUT, a dynamic detector allocation framework that improves prompt-injection defense by predicting detector reliability and latency to optimize the trade-off between safety and operational utility.