Refusal is Not Safety! Benchmarking Latent Safety Risks of LLM-Driven Content Humorization

2026-07-17

AI Summary

This study addresses a previously overlooked safety risk in large language models (LLMs): the use of humor as an indirect refusal strategy, which can implicitly introduce stereotypes and toxic content. Current safety evaluation frameworks lack coverage for such subtle hazards. The work presents the first systematic analysis demonstrating that humor itself can amplify safety risks, introducing HumorSafe—the first evaluation framework tailored to assess these implicit dangers—and HumorPIA, a novel prompt injection attack leveraging humor. Evaluated across five mainstream LLMs using over 30,000 real-world interactions, feedback from 45 professional comedians, and adversarial prompt engineering, the findings reveal that humor significantly exacerbates bias and toxicity. Notably, HumorPIA increases toxicity by 3.14× while maintaining a 97.8% apparent safety rate, effectively evading existing defense mechanisms.

Abstract

Safety defenses for large language models (LLMs) have been extensively studied, with existing approaches focusing on attack detection and refusal mechanisms. Such fixed-form direct refusal strategies may introduce the risk of prefix injection attacks. Recent work has explored a new direction that leverages humor as an indirect refusal mechanism to mitigate over-refusal in jailbreak scenarios and reduce prefix injection risks. However, this approach implicitly assumes that humorous responses are safe. Whether humorization itself introduces safety risks remains unexplored. To address this issue, we conduct an exploratory study involving over 30,000 real-world agent interaction records and 45 stand-up comedians, revealing practical safety concerns in LLM-based content humorization. Motivated by these findings, we propose \textsc{HumorSafe}, a novel framework for evaluating latent safety risk propagation during humorization. \textsc{HumorSafe} enables LLMs to learn harmful humorization patterns and use them to transform benign content into humorous content with safety risks. Across five frontier LLMs, we find that LLMs can introduce stereotypes and toxicity during humorization. We further propose \textsc{HumorPIA}, a prompt injection attack that exploits latent risks in humor-based defenses. \textsc{HumorPIA} preserves the appearance of safe humorous refusal while covertly injecting harmful content, allowing latent risks to evade existing detection mechanisms. Experiments show that it increases toxicity by 3.14$\times$ while maintaining an apparent safety rate of 97.8\% even under defense settings. Our findings highlight a gap in existing LLM safety evaluations under humorized settings.

Problem

• LLM safety
• humorization
• latent risks
• prompt injection
• toxicity

Innovation

• humorization
• latent safety risks
• prompt injection attack
• LLM safety evaluation
• indirect refusal