{"id":"speak-easy","title":"Speak Easy: Eliciting Harmful Jailbreaks from LLMs with Simple Interactions","authors":[{"given":"Yik Siu","family":"Chan"},{"given":"Narutatsu","family":"Ri"},{"given":"Yuxin","family":"Xiao"},{"given":"Marzyeh","family":"Ghassemi"}],"first_public":"2025-02-06","date_note":"First arXiv version.","date_source":"https://arxiv.org/abs/2502.04322","year":2025,"venue":"ICML 2025","paper_url":"https://proceedings.mlr.press/v267/chan25b.html","arxiv":"2502.04322","code_url":"https://github.com/yiksiu-chan/SpeakEasy","category":"safety","relationship":"uses","relationship_label":"Uses WildVis","dataset_version":"Real-world conversations retrieved through WildVis motivate the study. Experiments use HarmBench, AdvBench, SORRY-Bench, and MedSafetyBench.","summary":"Uses conversations found through WildVis to study safety failures during multi-step and multilingual interactions. The paper introduces HarmScore to assess the actionability and informativeness of harmful responses, and tests these interaction patterns across four safety benchmarks.","finding":"Figure 2 uses WildVis conversations to illustrate the interaction patterns tested in the safety evaluation.","evidence_url":"https://arxiv.org/html/2502.04322v3","evidence_note":"Section 4 and Figure 2: WildVis conversations. Section 5: benchmark evaluation.","bibtex_key":"chan2025speakeasy","record_url":"https://yuntiandeng.com/wildchat/papers/speak-easy/","citation_title":"Speak Easy: Eliciting Harmful Jailbreaks from LLMs with Simple Interactions","citation_type":"paper-conference","bibtex_type":"inproceedings","container_title":"Proceedings of the 42nd International Conference on Machine Learning","citation_url":"https://proceedings.mlr.press/v267/chan25b.html","citation_year":2025,"citation_note":"ICML 2025 proceedings paper, PMLR volume 267.","pages":"7306–7331","volume":"267","publisher":"PMLR"}