arxiv9.0 / 10

Question, Then Affirm: Restoring Commitment Does Not Restore Refusal

Haoyang Xu, Qirui Wang, Liuwan Zhu

Abstract

We find that safeguards for image generation can be bypassed without changing what a message says. A sign stating a harmful proposition is blocked, but the same proposition phrased as a question and followed by a check mark is generated, even though the resulting image endorses the original claim. On a 300-case challenge set, gpt-image-2 returns images for 43.3% of question-affirmation requests versus 16.3% of direct statements, with 41.3% versus 15.3% of cases yielding images that endorse the original proposition. We define this as Deferred Semantic Commitment (DSC): commitment is withdrawn by the question and restored by a separate element, while safety behavior follows only the withdrawal. To investigate why, we study an open-weight language model and find that re-endorsement only partially reverses the refusal-related change introduced by questioning. These findings reveal a mismatch between recovered meaning and refusal behavior, motivating safeguards that assess the completed message rather than its individual elements.

Research area

Adversarial evaluationjailbreakingmodel robustness
Published
19 Nov 2026
Source
arxiv
Org
University of Hawai’i at Manoa
View paper
Sign in to read and join the discussion.