Jiahang Wang, Chenglong Fu, Xiaojiang Du, Bo Luo · ACM SIGSAC Conference on Computer and Communications Security (CCS) 2026 · 2026
DOI: 10.1145/3830454.3846742
Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).
Aligned large language models (LLMs) often refuse harmful requests under direct prompting, motivating decomposition-based jailbreak attacks that split an unsafe task into multiple narrower sub-questions. Although such attacks can reduce refusal, existing methods typically rely on static or heuristic decomposition strategies and lack principled mechanisms for improving decomposition quality or selecting among decompositions of different granularities. In this work, we propose a unified decomposition-based jailbreak framework that combines reinforcement-learning-based policy improvement with neuron-guided adaptive selection. We first train a weak decomposer using online reward-based optimization, where generated sub-questions are evaluated by downstream refusal behavior and helpfulness. We then construct a neuron-based risk score based on internal activations of a strong model and use it at inference time to rank candidate decompositions and select the lowest-risk candidate for execution. The trained decomposer and reference-model scorer are reused across target models without target-specific training. Tested on multiple aligned strong models, the proposed pipeline consistently reduces refusal while maintaining helpfulness. Compared with representative automatic jailbreak baselines, it achieves stronger attack performance with substantially lower refusal exposure, which is particularly important in commercial API settings where repeated refusals may increase the risk of monitoring, throttling, or access restrictions. Our mechanistic analyses further suggest an activation dilution effect: when a harmful request is decomposed into separately executed sub-questions, safety-relevant signals are dispersed across queries, whereas recombining them into a single batched prompt partially restores risk. These results show that effective decomposition-based jailbreaks depend not only on whether a task is decomposed, but also on how the decomposition policy is learned and how decomposition granularity is controlled at inference time.
No comments yet — start the discussion below.