Wenzhuo Xu, Zhipeng Wei, Zonghao Ying, Deyue Zhang, Dongdong Yang, Xiangzheng Zhang, Quanchen Zou. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

The development of text-to-image (T2I) generative models, that enable the creation of high-quality synthetic images from textual prompts, has opened new frontiers in creative design and content generation. However, this paper reveals a significant and previously unrecognized ethical risk inherent in this technology and introduces a novel method, termed the Cognitive Morphing Attack(CogMorph), which manipulates T2I models to generate images that retain the original core subjects but embeds toxic or harmful contextual elements. This nuanced manipulation exploits the cognitive principle that human perception of concepts is shaped by the entire visual scene and its context, producing images that amplify emotional harm far beyond attacks that merely preserve the original semantics. To address this, we first construct an imagery toxicity taxonomy spanning 10 major and 48 sub-categories, aligned with human cognitive-perceptual dimensions, and further build a toxicity risk matrix resulting in 1,176 high-quality T2I toxic prompts. Based on this, ourCogMorphfirst introduces Cognitive Toxicity Augmentation, which develops a cognitive toxicity knowledge base with rich external toxic representations for humans (e.g., fine-grained visual features) that can be utilized to further guide the optimization of adversarial prompts. In addition, we present Contextual Hierarchical Morphing, which hierarchically extracts critical parts of the original prompt (e.g., scenes, subjects, and body parts), and then iteratively retrieves and fuses toxic features to inject harmful contexts. Extensive experiments on multiple open-source T2I models and black-box commercial APIs (e.g., DALL$\cdot$E-3) demonstrate the efficacy ofCogMorphwhich significantly outperforms other baselines by large margins (+20.62% on average). Our codes are available athttps://github.com/raykr/CogMorph.Warning: This paper contains harmful imagery that might be offensive to some readers.

The emergence of multimodal large language models has redefined the agent paradigm by integrating language and vision modalities with external data sources, enabling agents to better interpret human instructions and execute increasingly complex tasks. However, in this paper, we identify a critical yet previously overlooked security vulnerability in multimodal agents: cross-modal prompt injection attacks. To exploit this vulnerability, we propose CrossInject, a novel attack framework in which attacker embeds adversarial perturbations across multiple modalities to align with target malicious content, allowing external instructions to hijack the agents' decision-making process and execute unauthorized tasks. Our approach incorporates two key coordinated components. First, we introduce Visual Latent Alignment, where we optimize adversarial features to the malicious instructions in the visual embedding space based on a text-to-image generative model, ensuring that adversarial images subtly encode cues for malicious task execution. Subsequently, we present Textual Guidance Enhancement, where a large language model is leveraged to construct the black-box defensive system prompt through adversarial meta-prompting and generate a malicious textual command based on it that steers the agents' output toward better compliance with attacker's requests. Extensive experiments demonstrate that our method outperforms state-of-the-art attacks, achieving at least a +30.1% increase in attack success rates across diverse tasks. Furthermore, we validate our attack's effectiveness in real-world multimodal autonomous agents, highlighting its potential implications for safety-critical applications. Code can be found in https://github.com/Larry0454/CrossInject.

Large Language Models (LLMs) have demonstrated impressive performance across various tasks, yet they remain vulnerable to generating toxic content, necessitating detoxification strategies to ensure safe and responsible deployment. Test-time detoxification methods, which typically introduce static or dynamic interventions into LLM representations, offer a promising solution due to their flexibility and minimal invasiveness. However, current approaches often suffer from imprecise interventions, primarily due to their insufficient exploration of the transition space between toxic and non-toxic outputs. To address this challenge, we propose \textsc{A}utoregressive \textsc{R}eward \textsc{G}uided \textsc{R}epresentation \textsc{E}diting (ARGRE), a novel test-time detoxification framework that explicitly models toxicity transitions within the latent representation space, enabling stable and precise reward-guided editing. ARGRE identifies non-toxic semantic directions and interpolates between toxic and non-toxic representations to reveal fine-grained transition trajectories. These trajectories transform sparse toxicity annotations into dense training signals, enabling the construction of an autoregressive reward model that delivers stable and precise editing guidance. At inference, the reward model guides an adaptive two-step editing process to obtain detoxified representations: it first performs directional steering based on expected reward gaps to shift representations toward non-toxic regions, followed by lightweight gradient-based refinements. Extensive experiments across 8 widely used LLMs show that ARGRE significantly outperforms leading baselines in effectiveness (-62.21\% toxicity) and efficiency (-47.58\% inference time), while preserving the core capabilities of the original model with minimal degradation. Our code is available at the \href{https://anonymous.4open.science/r/ARGRE-6291}{anonymous website}.

In the realm of large vision language models (LVLMs), jailbreak attacks serve as a red-teaming approach to bypass guardrails and uncover safety implications. Existing jailbreaks predominantly focus on the visual modality, perturbing solely visual inputs in the prompt for attacks. However, they fall short when confronted with aligned models that fuse visual and textual features simultaneously for generation. To address this limitation, this paper introduces the Bi-Modal Adversarial Prompt Attack (BAP), which executes jailbreaks by optimizing textual and visual prompts cohesively. Initially, we adversarially embed universally adversarial perturbations in an image, guided by a few-shot query-agnostic corpus (e.g., affirmative prefixes and negative inhibitions). This process ensures that the adversarial image prompt LVLMs to respond positively to harmful queries. Subsequently, leveraging the image, we optimize textual prompts with specific harmful intent. In particular, we utilize a large language model to analyze jailbreak failures and employ chain-of-thought reasoning to refine textual prompts through a feedback-iteration manner. To validate the efficacy of our approach, we conducted extensive evaluations on various datasets and LVLMs, demonstrating that our BAP significantly outperforms other methods by large margins (+29.03% in attack success rate on average). Additionally, we showcase the potential of our attacks on black-box commercial LVLMs, such as GPT-4o and Gemini.

Abstract Deep learning models are well known to be susceptible to backdoor attack, where the attacker only needs to provide a tampered dataset on which the triggers are injected. Models trained on the dataset will passively implant the backdoor, and triggers on the input can mislead the models during testing. Our study shows that the model shows different learning behaviors in clean and poisoned subsets during training. Based on this observation, we propose a general training pipeline to defend against backdoor attacks actively. Benign models can be trained from the unreliable dataset by decoupling the learning process into three stages, i.e., supervised learning, active unlearning, and active semi-supervised fine-tuning. The effectiveness of our approach has been shown in numerous experiments across various backdoor attacks and datasets.

Abstract Recently, deep neural networks have been shown to be vulnerable to backdoor attacks. A backdoor is inserted into neural networks via this attack paradigm, thus compromising the integrity of the network. As soon as an attacker presents a trigger during the testing phase, the backdoor in the model is activated, allowing the network to make specific wrong predictions. It is extremely important to defend against backdoor attacks since they are very stealthy and dangerous. In this paper, we propose a novel defense mechanism, Neural Behavioral Alignment (NBA), for backdoor removal. NBA optimizes the distillation process in terms of knowledge form and distillation samples to improve defense performance according to the characteristics of backdoor defense. NBA builds high-level representations of neural behavior within networks in order to facilitate the transfer of knowledge. Additionally, NBA crafts pseudo samples to induce student models exhibit backdoor neural behavior. By aligning the backdoor neural behavior from the student network with the benign neural behavior from the teacher network, NBA enables the proactive removal of backdoors. Extensive experiments show that NBA can effectively defend against six different backdoor attacks and outperform five state-of-the-art defenses.

Security attacks in smart contracts have drawn extensive attention due to the financial loss and erosion of trust caused by vulnerabilities. Even worse, smart contract is a tamper proof digital agreement and fixing bugs in it is difficult, so it is necessary for developers to detect security vulnerabilities in smart contract before deployment. Researchers have proposed several methods on smart contract vulnerability detection. However, despite the numerous vulnerability-finding tools, few of them have ideal detection performance because most of them rely on fixed rules, which is inefficient. In this paper, a modularized and systematic Deep Learning-based framework is proposed to automatically detect smart contracts vulnerability, called DeeSCVHunter. Particularly, we focus on two types of smart contract vulnerabilities: reentrancy and time dependence. And we propose a novel notion of Vulnerability Candidate Slice (VCS) to help models capture the key point of vulnerability. We conduct experiments on real-world dataset and the results prove the effectiveness of VCS, which leads to 25.76% improvement in F1-score at most. And extensive experiments also show that our proposed framework significantly outperforms state-of-the-art methods.