Title: Examining and Mitigating Harm Amplification in Large Language Model Conversations

Large language models (LLMs) offer significant utility as assistants, but they also pose a risk of acting as force multipliers for malicious actors, enabling individuals to execute harmful actions that exceed their own skill sets through prolonged engagement. This threat operates along two primary dimensions: it democratizes specialized knowledge, allowing inexperienced users to generate expert-level harmful content, and it scales harmful activities to volumes unattainable through manual means. Despite these concerns, previous research has frequently failed to account for how LLMs exacerbate harm within multi-turn dialogue contexts.

To address this gap, we present HarmAmp, a comprehensive benchmark designed to evaluate multi-turn harm amplification across twelve distinct risk categories. Each scenario in this benchmark is rooted in actual real-world threats and adheres to strict standards, including substantive amplification, operational specificity, and the necessity of multiple interaction turns. Additionally, we introduce TrajSafe, a proactive monitoring system designed to predict dangerous conversation paths and intervene by probing users’ true intentions and guiding the model toward safer responses. Our extensive testing reveals that TrajSafe effectively lowers the level of harm generated during multi-turn exchanges, maintaining a low rate of unnecessary refusals and preserving the model’s overall general capabilities. This research provides a promising framework for addressing the subtle safety challenges inherent in LLM interactions.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC