Safety alignment in large language models (LLMs) is primarily evaluated under open-ended generation, where models can mitigate risk by refusing to respond. In contrast, many real-world applications place LLMs in structured decision-making tasks, such as multiple-choice questions (MCQs), where abstention is discouraged or unavailable. We identify a systematic failure mode in this setting: reformulating harmful requests as forced-choice MCQs, where all options are unsafe, can systematically bypass refusal behavior, even in models that consistently reject equivalent open-ended prompts. Across 14 proprietary and open-source models, we show that forced-choice constraints sharply increase policy-violating responses. Notably, for human-authored MCQs, violation rates follow an inverted U-shaped trend with respect to structural constraint strength, peaking under intermediate task specifications, whereas MCQs generated by high-capability models yield near-saturation violation rates across constraints and exhibit strong cross-model transferability. Our findings reveal that current safety evaluations substantially underestimate risks in structured task settings and highlight constrained decision-making as a critical and underexplored surface for alignment failures.

Deep neural network-based classifiers are prone to errors when processing adversarial examples (AEs). AEs are minimally perturbed input data undetectable to humans posing significant risks to security-dependent applications. Hence, extensive research has been undertaken to develop defense mechanisms that mitigate their threats. Most existing methods primarily focus on discriminating AEs based on the input sample features, emphasizing AE detection without addressing the correct sample categorization before an attack. While some tasks may only require mere rejection on detected AEs, others necessitate identifying the correct original input category such as traffic sign recognition in autonomous driving. The objective of this study is to propose a method for rectifying AEs to estimate the correct labels of their original inputs. Our method is based on re-attacking AEs to move them beyond the decision boundary for accurate label prediction, effectively addressing the issue of rectifying minimally perceptible AEs created using white-box attack methods. However, challenge remains with respect to effectively rectifying AEs produced by black-box attacks at a distance from the boundary, or those misclassified into low-confidence categories by targeted attacks. By adopting a straightforward approach of only considering AEs as inputs, the proposed method can address diverse attacks while avoiding the requirement of parameter adjustments or preliminary training. Results demonstrate that the proposed method exhibits consistent performance in rectifying AEs generated via various attack methods, including targeted and black-box attacks. Moreover, it outperforms conventional rectification and input transformation methods in terms of stability against various attacks.

Deep neural networks (DNNs) remain vulnerable to adversarial attacks that cause misclassification when specific perturbations are added to input images. This vulnerability also threatens the reliability of DNN-based monocular depth estimation (MDE) models, making robustness enhancement a critical need in practical applications. To validate the vulnerability of DNN-based MDE models, this study proposes a projection-based adversarial attack method that projects perturbation light onto a target object. The proposed method employs physics-in-the-loop (PITL) optimization -- evaluating candidate solutions in actual environments to account for device specifications and disturbances -- and utilizes a distributed covariance matrix adaptation evolution strategy. Experiments confirmed that the proposed method successfully created adversarial examples that lead to depth misestimations, resulting in parts of objects disappearing from the target scene.