Furina Attack Exploits LLM Safety Instability Region

A newly identified attack strategy, dubbed Furina, targets vulnerabilities in large language models (LLMs) and their multimodal counterparts. Researchers have highlighted that small changes in input can lead to erratic responses from these systems, revealing the complexity of safety alignment. Their findings, documented in a recent study on arXiv (2605.26158), suggest that current detection mechanisms struggle against advanced threats due to this unpredictability. The diagnostic tool they created assesses multiple signals, demonstrating that while unpredictable outputs arise in vulnerable areas, the system's internal safety responses remain compromised. Notably, Furina can be executed using segmented, scene-oriented prompts without altering the models directly.