Intent-Hiding Jailbreaks: An Information-Theoretic Framework for Compositional Attacks

2026-10-01

AI Summary

This work addresses the vulnerability of large language models to compositional jailbreak attacks, wherein harmful intents are concealed by concatenating benign tasks to evade safety detection. From an information-theoretic perspective, we propose a prior-posterior matching framework based on probability estimation that achieves intent concealment while preserving target behaviors. We further derive the optimal water-filling solution under score weighting and establish a minimum beam size threshold. Multi-model experiments demonstrate that this compositional query strategy significantly outperforms direct request baselines and reveals a theoretical trade-off between beam size and response-level goal preservation rates.

Abstract

Recent work has shown that large language models (LLMs) can be vulnerable to jailbreak attacks in which harmful intent is obscured through composition with benign tasks. A harmful request refused in isolation may elicit a different response when embedded within a larger, seemingly benign query. We study these compositional intent-hiding jailbreaks from an information-theoretic perspective. Our formulation associates each task with an estimated probability of being judged harmful: the average over the full task collection defines the prior probability of harmful intent, while the average over a selected bundle containing the target defines the posterior. Selecting auxiliary tasks so that these averages agree, which we call prior-posterior matching, leaves the estimated intent unchanged even though the harmful target remains in the bundle. We study two settings that differ in whether query construction is part of the optimization. In the query-independent setting, tasks are selected without regard to how they will be expressed in the final query. We show that exact prior-posterior matching under a bundle-size constraint is computationally hard, derive an optimal water-filling solution for fractional weights, and characterize the smallest bundle satisfying a prescribed safety threshold. In the query-dependent setting, task selection and query construction are considered jointly, and intent concealment and target preservation are evaluated on the resulting query. We evaluate jailbreak effectiveness and preservation of the target behavior across bundle sizes, query generators, and several open-source models. These results show that compositional queries can elicit target behaviors beyond the direct-request baseline under the evaluated search budgets, while revealing a trade-off: as bundle size increases, response-level target preservation tends to decrease for several models.

Problem

• jailbreak attacks
• large language models
• intent hiding
• compositional attacks
• information theory

Innovation

• information-theoretic framework
• prior-posterior matching
• compositional jailbreaks
• water-filling solution
• intent hiding