Researchers have introduced ReasAlign, a model-level solution designed to improve safety alignment in Large Language Models (LLMs) against indirect prompt injection attacks. The approach incorporates structured reasoning steps to analyze user queries and detect conflicting instructions, ensuring the continuity of the user's intended tasks.

  • ReasAlign employs a test-time scaling mechanism with a preference-optimized judge model to score reasoning steps and select the best trajectory.
  • On the CyberSecEval2 benchmark, it achieves 94.6% utility and only 3.6% attack success rate (ASR).
  • This performance far surpasses Meta SecAlign, which recorded 56.4% utility and 74.4% ASR on the same tasks.

The authors state that ReasAlign establishes a robust and practical defense against prompt injection attacks in real-world agentic systems by achieving the best trade-off between security and utility.