InDe-LLM: Defending against Jailbreak Attacks in LLM-Powered Systems via Intention Disentangling
Proceedings of the ACM on Software Engineering · 30.06.2026 · 3 (FSE) · 899-919
DOI: 10.1145/3797136 ↗Kurz gefasst
Fachartikel in der Zeitschrift Proceedings of the ACM on Software Engineering, online erschienen am 30.06.2026. Verfasst von Yujue Wang und 5 weiteren Autor:innen. Die Arbeit nennt 55 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY-NC-ND 4.0).
Abstract (Originalsprache)
Jailbreak attacks have been regarded as a crucial threat to LLM-powered software systems. Recent studies indicate the existence of a steering vector within models' internal activations, which can adjust a model's propensity to reject user requests, and thus is regarded as an effective approach for training-free defense. However, attackers may wrap their malicious intentions within a seemingly benign context, which shifts the distribution of harmful prompts toward benign inputs along the steering vector, effectively bypassing existing defense approaches. In this work, we propose a defense framework InDe-LLM based on intention disentangling. By projecting the embedding of inputs into a benign-invariant subspace, we could disentangle the harmful intentions of jailbreak prompts without affecting benign inputs. Next, such disentangled harmful intentions can be easily identified based on LLMs' well-aligned concept of harmfulness, and rejected through activation steering. Our experiments show that InDe-LLM achieves high defense effectiveness, outperforming baselines by 27.2%–43.5% across three models and ten attacks while preserving high utility on benign inputs. Moreover, our evaluation demonstrates that it exhibits high transferability to unseen attacks.
Zitieren
APA
Wang, Y., Zhang, Q., Zhou, C., Go, G., Shi, D., & Jiang, Y. (2026). InDe-LLM: Defending against Jailbreak Attacks in LLM-Powered Systems via Intention Disentangling. Proceedings of the ACM on Software Engineering, 3(FSE), 899-919. https://doi.org/10.1145/3797136
Vancouver
Wang Y, Zhang Q, Zhou C, Go G, Shi D, Jiang Y. InDe-LLM: Defending against Jailbreak Attacks in LLM-Powered Systems via Intention Disentangling. Proceedings of the ACM on Software Engineering. 2026;3(FSE):899-919. doi:10.1145/3797136
BibTeX
@article{Wang2026,
author = {Wang, Yujue and Zhang, Quan and Zhou, Chijin and Go, Gwihwan and Shi, Dalong and Jiang, Yu},
title = {InDe-LLM: Defending against Jailbreak Attacks in LLM-Powered Systems via Intention Disentangling},
journal = {Proceedings of the ACM on Software Engineering},
year = {2026},
volume = {3},
number = {FSE},
pages = {899--919},
doi = {10.1145/3797136}
}
EUVIMED