Emerging from Ground: Addressing Intent Deviation in Tool-Using Agents via Deriving Real Calls into Virtual Trajectories

LLMs have advanced tool-using agents for real-world applications, yet they often lead to unexpected behaviors or results. Beyond obvious failures, the subtle issue of "intent deviation" severely hinders reliable evaluation and performance improvement. Existing post-training methods generally leverage either real system samples or virtual data simulated by LLMs. However, the former is costly due to reliance on hand-crafted user requests, while the latter suffers from distribution shift from the real tools in the wild. Additionally, both methods lack negative samples tailored to intent deviation scenarios, hindering effective guidance on preference learning. We introduce RISE, a "Real-to-Virtual" method designed to mitigate intent deviation. Anchoring on verified tool primitives, RISE synthesizes virtual trajectories and generates diverse negative samples through mutation on critical parameters. With synthetic data, RISE fine-tunes backbone LLMs via the two-stage training for intent alignment. Evaluation results demonstrate that data synthesized by RISE achieve promising results in eight metrics covering user requires, execution trajectories and agent responses. Integrating with training, RISE achieves an average 35.28% improvement in Acctask (task completion) and 23.27% in Accintent (intent alignment), outperforming SOTA baselines by 1.20--42.09% and 1.17--54.93% respectively.

Know Thy Enemy: Securing LLMs Against Prompt Injection via Diverse Data Synthesis and Instruction-Level Chain-of-Thought Learning

Large language model (LLM)-integrated applications have become increasingly prevalent, yet face critical security vulnerabilities from prompt injection (PI) attacks. Defending against PI attacks faces two major issues: malicious instructions can be injected through diverse vectors, and injected instructions often lack clear semantic boundaries from the surrounding context, making them difficult to identify. To address these issues, we propose InstruCoT, a model enhancement method for PI defense that synthesizes diverse training data and employs instruction-level chain-of-thought fine-tuning, enabling LLMs to effectively identify and reject malicious instructions regardless of their source or position in the context. We evaluate InstruCoT across three critical dimensions: Behavior Deviation, Privacy Leakage, and Harmful Output. Experimental results across four LLMs demonstrate that InstruCoT significantly outperforms baselines in all dimensions while maintaining utility performance without degradation

All Changes May Have Invariant Principles: Improving Ever-Shifting Harmful Meme Detection via Design Concept Reproduction

Harmful memes are ever-shifting in the Internet communities, which are difficult to analyze due to their type-shifting and temporal-evolving nature. Although these memes are shifting, we find that different memes may share invariant principles, i.e., the underlying design concept of malicious users, which can help us analyze why these memes are harmful. In this paper, we propose RepMD, an ever-shifting harmful meme detection method based on the design concept reproduction. We first refer to the attack tree to define the Design Concept Graph (DCG), which describes steps that people may take to design a harmful meme. Then, we derive the DCG from historical memes with design step reproduction and graph pruning. Finally, we use DCG to guide the Multimodal Large Language Model (MLLM) to detect harmful memes. The evaluation results show that RepMD achieves the highest accuracy with 81.1% and has slight accuracy decreases when generalized to type-shifting and temporal-evolving memes. Human evaluation shows that RepMD can improve the efficiency of human discovery on harmful memes, with 15$\sim$30 seconds per meme.

One Shot Dominance: Knowledge Poisoning Attack on Retrieval-Augmented Generation Systems

Large Language Models (LLMs) enhanced with Retrieval-Augmented Generation (RAG) have shown improved performance in generating accurate responses. However, the dependence on external knowledge bases introduces potential security vulnerabilities, particularly when these knowledge bases are publicly accessible and modifiable. Poisoning attacks on knowledge bases for RAG systems face two fundamental challenges: the injected malicious content must compete with multiple authentic documents retrieved by the retriever, and LLMs tend to trust retrieved information that aligns with their internal memorized knowledge. Previous works attempt to address these challenges by injecting multiple malicious documents, but such saturation attacks are easily detectable and impractical in real-world scenarios. To enable the effective single document poisoning attack, we propose AuthChain, a novel knowledge poisoning attack method that leverages Chain-of-Evidence theory and authority effect to craft more convincing poisoned documents. AuthChain generates poisoned content that establishes strong evidence chains and incorporates authoritative statements, effectively overcoming the interference from both authentic documents and LLMs' internal knowledge. Extensive experiments across six popular LLMs demonstrate that AuthChain achieves significantly higher attack success rates while maintaining superior stealthiness against RAG defense mechanisms compared to state-of-the-art baselines.

Mimicking the Familiar: Dynamic Command Generation for Information Theft Attacks in LLM Tool-Learning System

Information theft attacks pose a significant risk to Large Language Model (LLM) tool-learning systems. Adversaries can inject malicious commands through compromised tools, manipulating LLMs to send sensitive information to these tools, which leads to potential privacy breaches. However, existing attack approaches are black-box oriented and rely on static commands that cannot adapt flexibly to the changes in user queries and the invocation chain of tools. It makes malicious commands more likely to be detected by LLM and leads to attack failure. In this paper, we propose AutoCMD, a dynamic attack comment generation approach for information theft attacks in LLM tool-learning systems. Inspired by the concept of mimicking the familiar, AutoCMD is capable of inferring the information utilized by upstream tools in the toolchain through learning on open-source systems and reinforcement with target system examples, thereby generating more targeted commands for information theft. The evaluation results show that AutoCMD outperforms the baselines with +13.2% $ASR_{Theft}$, and can be generalized to new tool-learning systems to expose their information leakage risks. We also design four defense methods to effectively protect tool-learning systems from the attack.

What External Knowledge is Preferred by LLMs? Characterizing and Exploring Chain of Evidence in Imperfect Context

Incorporating external knowledge into large language models (LLMs) has emerged as a promising approach to mitigate outdated knowledge and hallucination in LLMs. However, external knowledge is often imperfect. In addition to useful knowledge, external knowledge is rich in irrelevant or misinformation in the context that can impair the reliability of LLM responses. This paper focuses on LLMs' preferred external knowledge in imperfect contexts when handling multi-hop QA. Inspired by criminal procedural law's Chain of Evidence (CoE), we characterize that knowledge preferred by LLMs should maintain both relevance to the question and mutual support among knowledge pieces. Accordingly, we propose an automated CoE discrimination approach and explore LLMs' preferences from their effectiveness, faithfulness and robustness, as well as CoE's usability in a naive Retrieval-Augmented Generation (RAG) case. The evaluation on five LLMs reveals that CoE enhances LLMs through more accurate generation, stronger answer faithfulness, better robustness against knowledge conflict, and improved performance in a popular RAG case.

Repairing Catastrophic-Neglect in Text-to-Image Diffusion Models via Attention-Guided Feature Enhancement

Text-to-Image Diffusion Models (T2I DMs) have garnered significant attention for their ability to generate high-quality images from textual descriptions. However, these models often produce images that do not fully align with the input prompts, resulting in semantic inconsistencies. The most prominent issue among these semantic inconsistencies is catastrophic-neglect, where the images generated by T2I DMs miss key objects mentioned in the prompt. We first conduct an empirical study on this issue, exploring the prevalence of catastrophic-neglect, potential mitigation strategies with feature enhancement, and the insights gained. Guided by the empirical findings, we propose an automated repair approach named Patcher to address catastrophic-neglect in T2I DMs. Specifically, Patcher first determines whether there are any neglected objects in the prompt, and then applies attention-guided feature enhancement to these neglected objects, resulting in a repaired prompt. Experimental results on three versions of Stable Diffusion demonstrate that Patcher effectively repairs the issue of catastrophic-neglect, achieving 10.1%-16.3% higher Correct Rate in image generation compared to baselines.

VEglue: Testing Visual Entailment Systems via Object-Aligned Joint Erasing

Visual entailment (VE) is a multimodal reasoning task consisting of image-sentence pairs whereby a promise is defined by an image, and a hypothesis is described by a sentence. The goal is to predict whether the image semantically entails the sentence. VE systems have been widely adopted in many downstream tasks. Metamorphic testing is the commonest technique for AI algorithms, but it poses a significant challenge for VE testing. They either only consider perturbations on single modality which would result in ineffective tests due to the destruction of the relationship of image-text pair, or just conduct shallow perturbations on the inputs which can hardly detect the decision error made by VE systems. Motivated by the fact that objects in the image are the fundamental element for reasoning, we propose VEglue, an object-aligned joint erasing approach for VE systems testing. It first aligns the object regions in the premise and object descriptions in the hypothesis to identify linked and un-linked objects. Then, based on the alignment information, three Metamorphic Relations are designed to jointly erase the objects of the two modalities. We evaluate VEglue on four widely-used VE systems involving two public datasets. Results show that VEglue could detect 11,609 issues on average, which is 194%-2,846% more than the baselines. In addition, VEglue could reach 52.5% Issue Finding Rate (IFR) on average, and significantly outperform the baselines by 17.1%-38.2%. Furthermore, we leverage the tests generated by VEglue to retrain the VE systems, which largely improves model performance (50.8% increase in accuracy) on newly generated tests without sacrificing the accuracy on the original test set.

Adversarial Testing for Visual Grounding via Image-Aware Property Reduction

Due to the advantages of fusing information from various modalities, multimodal learning is gaining increasing attention. Being a fundamental task of multimodal learning, Visual Grounding (VG), aims to locate objects in images through natural language expressions. Ensuring the quality of VG models presents significant challenges due to the complex nature of the task. In the black box scenario, existing adversarial testing techniques often fail to fully exploit the potential of both modalities of information. They typically apply perturbations based solely on either the image or text information, disregarding the crucial correlation between the two modalities, which would lead to failures in test oracles or an inability to effectively challenge VG models. To this end, we propose PEELING, a text perturbation approach via image-aware property reduction for adversarial testing of the VG model. The core idea is to reduce the property-related information in the original expression meanwhile ensuring the reduced expression can still uniquely describe the original object in the image. To achieve this, PEELING first conducts the object and properties extraction and recombination to generate candidate property reduction expressions. It then selects the satisfied expressions that accurately describe the original object while ensuring no other objects in the image fulfill the expression, through querying the image with a visual understanding technique. We evaluate PEELING on the state-of-the-art VG model, i.e. OFA-VG, involving three commonly used datasets. Results show that the adversarial tests generated by PEELING achieves 21.4% in MultiModal Impact score (MMI), and outperforms state-of-the-art baselines for images and texts by 8.2%--15.1%.

Play Guessing Game with LLM: Indirect Jailbreak Attack with Implicit Clues

With the development of LLMs, the security threats of LLMs are getting more and more attention. Numerous jailbreak attacks have been proposed to assess the security defense of LLMs. Current jailbreak attacks primarily utilize scenario camouflage techniques. However their explicitly mention of malicious intent will be easily recognized and defended by LLMs. In this paper, we propose an indirect jailbreak attack approach, Puzzler, which can bypass the LLM's defense strategy and obtain malicious response by implicitly providing LLMs with some clues about the original malicious query. In addition, inspired by the wisdom of "When unable to attack, defend" from Sun Tzu's Art of War, we adopt a defensive stance to gather clues about the original malicious query through LLMs. Extensive experimental results show that Puzzler achieves a query success rate of 96.6% on closed-source LLMs, which is 57.9%-82.7% higher than baselines. Furthermore, when tested against the state-of-the-art jailbreak detection approaches, Puzzler proves to be more effective at evading detection compared to baselines.