KnowJail: Knowledge Enhanced Large Language Model Jailbreak

摘要

Large language models (LLMs) are increasingly deployed in specialized domains, where unsafe behavior may arise from professional concepts, relations, and scenarios that are underrepresented in general-purpose safety benchmarks. Existing jailbreak evaluations are mostly prompt-centric, revealing whether specific prompts bypass safety mechanisms but not which domain knowledge units give rise to unsafe behavior. We propose KnowJail, a knowledge-grounded framework that reformulates domain-specific jailbreak discovery as knowledge-conditioned vulnerability estimation. Given domain knowledge text and black-box access to a target model, KnowJail organizes knowledge into lightweight triples, derives risk-oriented candidates, and estimates their vulnerability through offline surrogate exploration and limited online refinement. This enables knowledge-level risk attribution, allowing evaluators to identify vulnerable domain relations rather than only successful jailbreak prompts. Experiments across multiple domains and target models show that KnowJail consistently uncovers domain-specific vulnerabilities and improves attack effectiveness under limited query budgets. The code is available at \url{https://anonymous.4open.science/status/kg_jail}.