{"status":"ok","message-type":"work","message-version":"1.0.0","message":{"indexed":{"date-parts":[[2026,9,21]],"date-time":"2026-09-21T17:11:17Z","timestamp":1790010677676,"version":"4.0.1"},"reference-count":37,"publisher":"MIT Press","license":[{"start":{"date-parts":[[2026,9,21]],"date-time":"2026-09-21T00:00:00Z","timestamp":1789948800000},"content-version":"vor","delay-in-days":263,"URL":"https:\/\/creativecommons.org\/licenses\/by\/4.0\/"}],"content-domain":{"domain":["direct.mit.edu"],"crossmark-restriction":true},"short-container-title":[],"published-print":{"date-parts":[[2026,9,18]]},"abstract":"<jats:title>Abstract<\/jats:title>\n                  <jats:p>Large language models (LLMs) remain vulnerable to jailbreaking attacks despite their impressive capabilities. Investigating these vulnerabilities is crucial for developing robust safety mechanisms. Existing attacks primarily bypass LLM safeguards by introducing additional context or adversarial tokens, while leaving the core harmful intent intact. In this paper, we introduce ISA (Intent Shift Attack), which conceals malicious intent from LLMs through subtle edits. More specifically, we establish a taxonomy of intent transformations and leverage them to generate attacks that may be misperceived by LLMs as benign requests. Unlike prior methods relying on complex tokens or lengthy context, our approach requires only limited modifications to the original request, yielding natural, fluent, and seemingly harmless prompts. Extensive experiments on both open-source and commercial LLMs demonstrate that ISA effectively induces safety failures, achieving high attack success rates through simple linguistic transformations. More critically, fine-tuning models solely on benign data reformulated with ISA templates elevates the success rates to nearly 100%. For defense, we evaluate existing methods and demonstrate their limited effectiveness against ISA, while exploring both training-free and training-based mitigation strategies. Our findings reveal fundamental challenges in intent inference for LLM safety and underscore the need for more effective defenses. Our code and datasets are available at https:\/\/github.com\/NJUNLP\/ISA.<\/jats:p>","DOI":"10.1162\/tacl.a.805","type":"journal-article","created":{"date-parts":[[2026,9,21]],"date-time":"2026-09-21T16:34:49Z","timestamp":1790008489000},"page":"2357-2373","update-policy":"https:\/\/doi.org\/10.1162\/mitpressjournals.corrections.policy","source":"Crossref","is-referenced-by-count":0,"title":["Friend or Foe: How LLMs\u2019 Safety Mind Gets Fooled by Intent Shift Attack"],"prefix":"10.1162","volume":"14","author":[{"given":"Peng","family":"Ding","sequence":"first","affiliation":[{"name":"National Key Laboratory for Novel Software Technology, Nanjing University, dingpeng@smail.nju.edu.cn"}],"role":[{"vocabulary":"crossref","role":"author"}]},{"given":"Jun","family":"Kuang","sequence":"additional","affiliation":[{"name":"Meituan Inc., China, kuangjun@meituan.com"}],"role":[{"vocabulary":"crossref","role":"author"}]},{"given":"Wen","family":"Sun","sequence":"additional","affiliation":[{"name":"Meituan Inc., China, sunwen16@meituan.com"}],"role":[{"vocabulary":"crossref","role":"author"}]},{"given":"Zongyu","family":"Wang","sequence":"additional","affiliation":[{"name":"Meituan Inc., China, wangzongyu02@meituan.com"}],"role":[{"vocabulary":"crossref","role":"author"}]},{"given":"Xuezhi","family":"Cao","sequence":"additional","affiliation":[{"name":"Meituan Inc., China, caoxuezhi@meituan.com"}],"role":[{"vocabulary":"crossref","role":"author"}]},{"given":"Xunliang","family":"Cai","sequence":"additional","affiliation":[{"name":"Meituan Inc., China, caixunliang@meituan.com"}],"role":[{"vocabulary":"crossref","role":"author"}]},{"given":"Jiajun","family":"Chen","sequence":"additional","affiliation":[{"name":"National Key Laboratory for Novel Software Technology, Nanjing University, chenjj@nju.edu.cn"}],"role":[{"vocabulary":"crossref","role":"author"}]},{"given":"Shujian","family":"Huang","sequence":"additional","affiliation":[{"name":"National Key Laboratory for Novel Software Technology, Nanjing University, huangsj@nju.edu.cn"}],"role":[{"vocabulary":"crossref","role":"author"}]}],"member":"281","published-online":{"date-parts":[[2026,9,18]]},"reference":[{"key":"2026092112343723800_bib1","article-title":"Does refusal training in LLMs generalize to the past tense?","volume-title":"The Thirteenth International Conference on Learning Representations, ICLR 2025, Singapore, April 24-28, 2025","author":"Andriushchenko","year":"2025"},{"key":"2026092112343723800_bib2","doi-asserted-by":"crossref","DOI":"10.52202\/079017-1745","article-title":"Jailbreakbench: An open robustness benchmark for jailbreaking large language models","volume-title":"Advances in Neural Information Processing Systems 37: Annual Conference on Neural Information Processing Systems 2024, NeurIPS 2024, Vancouver, BC, Canada, December 10 - 15, 2024","author":"Chao","year":"2024"},{"key":"2026092112343723800_bib3","first-page":"23","article-title":"Jailbreaking black box large language models in twenty queries","volume-title":"IEEE Conference on Secure and Trustworthy Machine Learning, SaTML 2025, Copenhagen, Denmark, April 9-11, 2025","author":"Chao","year":"2025"},{"key":"2026092112343723800_bib4","first-page":"4299","article-title":"Deep reinforcement learning from human preferences","volume-title":"Advances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems 2017, December 4-9, 2017, Long Beach, CA, USA","author":"Christiano","year":"2017"},{"key":"2026092112343723800_bib5","article-title":"Deepseek-r1: Incentivizing reasoning capability in LLMs via reinforcement learning","volume-title":"CoRR","author":"DeepSeek-AI","year":"2025"},{"key":"2026092112343723800_bib6","first-page":"2136","article-title":"A wolf in sheep\u2019s clothing: Generalized nested jailbreak prompts can fool large language models easily","volume-title":"Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), NAACL 2024, Mexico City, Mexico, June 16-21, 2024","author":"Ding","year":"2024"},{"key":"2026092112343723800_bib7","first-page":"6279","article-title":"Why not act on what you know? unleashing safety potential of LLMs via self-aware guard enhancement","volume-title":"Findings of the Association for Computational Linguistics, ACL 2025, Vienna, Austria, July 27 - August 1, 2025","author":"Ding","year":"2025"},{"key":"2026092112343723800_bib8","first-page":"5023","article-title":"SDGO: self-discrimination-guided optimization for consistent safety in large language models","volume-title":"Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, EMNLP 2025, Suzhou, China, November 4-9, 2025","author":"Ding","year":"2025"},{"key":"2026092112343723800_bib9","first-page":"6734","article-title":"Attacks, defenses and evaluations for LLM conversation safety: A survey","volume-title":"Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), NAACL 2024, Mexico City, Mexico, June 16-21, 2024","author":"Dong","year":"2024"},{"key":"2026092112343723800_bib10","article-title":"The ethics of advanced AI assistants","volume-title":"CoRR","author":"Gabriel","year":"2024"},{"key":"2026092112343723800_bib11","article-title":"What is in your safe data? identifying benign data that breaks safety","volume-title":"CoRR","author":"He","year":"2024"},{"key":"2026092112343723800_bib12","article-title":"Measuring massive multitask language understanding","volume-title":"9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021","author":"Hendrycks","year":"2021"},{"key":"2026092112343723800_bib13","article-title":"Lora: Lowrank adaptation of large language models","volume-title":"The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022","author":"Hu","year":"2022"},{"key":"2026092112343723800_bib14","article-title":"Catastrophic jailbreak of open-source LLMs via exploiting generation","volume-title":"The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024","author":"Huang","year":"2024"},{"key":"2026092112343723800_bib15","article-title":"Baseline defenses for adversarial attacks against aligned language models","volume-title":"CoRR","author":"Jain","year":"2023"},{"key":"2026092112343723800_bib16","article-title":"Deepinception: Hypnotize large language model to be jailbreaker","volume-title":"CoRR","author":"Li","year":"2023"},{"key":"2026092112343723800_bib17","volume-title":"Alpacaeval: An automatic evaluator of instruction-following models","author":"Li","year":"2023"},{"key":"2026092112343723800_bib18","article-title":"Autodan: Generating stealthy jailbreak prompts on aligned large language models","volume-title":"The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024","author":"Liu","year":"2024"},{"key":"2026092112343723800_bib19","article-title":"Jailbreaking chatgpt via prompt engineering: An empirical study","volume-title":"CoRR","author":"Liu","year":"2023"},{"key":"2026092112343723800_bib20","article-title":"GPT-4 technical report","volume-title":"CoRR","author":"OpenAI","year":"2023"},{"key":"2026092112343723800_bib21","doi-asserted-by":"crossref","DOI":"10.52202\/068431-2011","article-title":"Training language models to follow instructions with human feedback","volume-title":"Advances in Neural Information Processing Systems 35: Annual Conference on Neural Information Processing Systems 2022, NeurIPS 2022, New Orleans, LA, USA, November 28 - December 9, 2022","author":"Ouyang","year":"2022"},{"key":"2026092112343723800_bib22","article-title":"LLM self defense: By self examination, LLMs know they are being tricked","volume-title":"The Second Tiny Papers Track at ICLR 2024, Tiny Papers @ ICLR 2024, Vienna, Austria, May 11, 2024","author":"Phute","year":"2024"},{"key":"2026092112343723800_bib23","article-title":"Fine-tuning aligned language models compromises safety, even when users do not intend to!","volume-title":"The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024","author":"Qi","year":"2024"},{"key":"2026092112343723800_bib24","first-page":"11437","article-title":"Codeattack: Revealing safety generalization challenges of large language models via code completion","volume-title":"Findings of the Association for Computational Linguistics, ACL 2024, Bangkok, Thailand and virtual meeting, August 11-16, 2024","author":"Ren","year":"2024"},{"key":"2026092112343723800_bib25","first-page":"5377","article-title":"Xstest: A test suite for identifying exaggerated safety behaviours in large language models","volume-title":"Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), NAACL 2024, Mexico City, Mexico, June 16-21, 2024","author":"R\u00f6ttger","year":"2024"},{"key":"2026092112343723800_bib26","first-page":"1671","article-title":"\u201cDo anything now\u201d: Characterizing and evaluating in-the-wild jailbreak prompts on large language models","volume-title":"Proceedings of the 2024 on ACM SIGSAC Conference on Computer and Communications Security, CCS 2024, Salt Lake City, UT, USA, October 14-18, 2024","author":"Shen","year":"2024"},{"key":"2026092112343723800_bib27","article-title":"The Llama 3 herd of models","volume-title":"CoRR","author":"Team","year":"2024"},{"key":"2026092112343723800_bib28","doi-asserted-by":"crossref","first-page":"6835","DOI":"10.1109\/TPAMI.2026.3660147","article-title":"Jailbreak and guard aligned language models with only few in-context demonstrations","volume":"48","author":"Wei","year":"2026","journal-title":"IEEE Transactions on Pattern Analysis and Machine Intelligence"},{"issue":"12","key":"2026092112343723800_bib29","doi-asserted-by":"crossref","first-page":"1486","DOI":"10.1038\/s42256-023-00765-8","article-title":"Defending chatgpt against jailbreak attack via self-reminders","volume":"5","author":"Xie","year":"2023","journal-title":"Nat. Mac. Intell."},{"key":"2026092112343723800_bib30","first-page":"5587","article-title":"Safedecoding: Defending against jailbreak attacks via safety-aware decoding","volume-title":"Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), ACL 2024, Bangkok, Thailand, August 11-16, 2024","author":"Xu","year":"2024"},{"key":"2026092112343723800_bib31","article-title":"Qwen2.5 technical report","volume-title":"CoRR","author":"Yang","year":"2024"},{"key":"2026092112343723800_bib32","article-title":"GPTFUZZER: red teaming large language models with auto-generated jailbreak prompts","volume-title":"CoRR","author":"Yu","year":"2023"},{"key":"2026092112343723800_bib33","first-page":"14322","article-title":"How Johnny can persuade llms to jailbreak them: Rethinking persuasion to challenge AI safety by humanizing LLMs","volume-title":"Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), ACL 2024, Bangkok, Thailand, August 11-16, 2024","author":"Zeng","year":"2024"},{"key":"2026092112343723800_bib34","first-page":"2947","article-title":"Intention analysis makes LLMs a good jailbreak defender","volume-title":"Proceedings of the 31st International Conference on Computational Linguistics, COLING 2025, Abu Dhabi, UAE, January 19-24, 2025","author":"Zhang","year":"2025"},{"key":"2026092112343723800_bib35","first-page":"8865","article-title":"Defending large language models against jail-breaking attacks through goal prioritization","volume-title":"Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), ACL 2024, Bangkok, Thailand, August 11-16, 2024","author":"Zhang","year":"2024"},{"key":"2026092112343723800_bib36","first-page":"61593","article-title":"On prompt-driven safeguarding for large language models","volume-title":"Forty-first International Conference on Machine Learning, ICML 2024, Vienna, Austria, July 21-27, 2024","author":"Zheng","year":"2024"},{"key":"2026092112343723800_bib37","article-title":"Universal and transferable adversarial attacks on aligned language models","volume-title":"CoRR","author":"Zou","year":"2023"}],"container-title":["Transactions of the Association for Computational Linguistics"],"original-title":[],"language":"en","link":[{"URL":"https:\/\/direct.mit.edu\/tacl\/article-pdf\/doi\/10.1162\/TACL.a.805\/2629462\/tacl.a.805.pdf","content-type":"application\/pdf","content-version":"vor","intended-application":"syndication"},{"URL":"https:\/\/direct.mit.edu\/tacl\/article-pdf\/doi\/10.1162\/TACL.a.805\/2629462\/tacl.a.805.pdf","content-type":"unspecified","content-version":"vor","intended-application":"similarity-checking"}],"deposited":{"date-parts":[[2026,9,21]],"date-time":"2026-09-21T16:34:58Z","timestamp":1790008498000},"score":1,"resource":{"primary":{"URL":"https:\/\/direct.mit.edu\/tacl\/article\/doi\/10.1162\/TACL.a.805\/139076\/Friend-or-Foe-How-LLMs-Safety-Mind-Gets-Fooled-by"}},"subtitle":[],"short-title":[],"issued":{"date-parts":[[2026]]},"references-count":37,"URL":"https:\/\/doi.org\/10.1162\/tacl.a.805","relation":{},"ISSN":["2307-387X"],"issn-type":[{"value":"2307-387X","type":"electronic"}],"subject":[],"published-other":{"date-parts":[[2026]]},"published":{"date-parts":[[2026]]}}}