JailSmash: Lightweight Jailbreaking of LLM Applications on Android Smartphones

The rapid growth of LLM APPs brings potential security risks, particularly jailbreaking attacks. These attacks involve bypassing LLM security mechanisms through carefully designed prompts to generate harmful or unethical content, posing threats to both individuals and organizations. We developed a lightweight jailbreak tool called JailSmash, which leverages jailbreak templates based on in-context learning (ICL) and strategies to guide third-party large language models in automatically generating jailbreak test sets. This tool also enables automated interaction and evaluation of jailbreak results with LLM apps on Android smartphones. Furthermore, we used JailSmash to conduct jailbreak security assessments on 30 mainstream LLM apps, thereby validating the effectiveness of the tool.