Lihai Nie

Biao Yi, Tiansheng Huang, Baolei Zhang et al.

Fine-tuning-as-a-service, while commercially successful for Large Language Model (LLM) providers, exposes models to harmful fine-tuning attacks. As a widely explored defense paradigm against such attacks, unlearning attempts to remove malicious knowledge from LLMs, thereby essentially preventing them from being used to perform malicious tasks. However, we highlight a critical flaw: the powerful general adaptability of LLMs allows them to easily bypass selective unlearning by rapidly relearning or repurposing their capabilities for harmful tasks. To address this fundamental limitation, we propose a paradigm shift: instead of selective removal, we advocate for inducing model collapse--effectively forcing the model to "unlearn everything"--specifically in response to updates characteristic of malicious adaptation. This collapse directly neutralizes the very general capabilities that attackers exploit, tackling the core issue unaddressed by selective unlearning. We introduce the Collapse Trap (CTRAP) as a practical mechanism to implement this concept conditionally. Embedded during alignment, CTRAP pre-configures the model's reaction to subsequent fine-tuning dynamics. If updates during fine-tuning constitute a persistent attempt to reverse safety alignment, the pre-configured trap triggers a progressive degradation of the model's core language modeling abilities, ultimately rendering it inert and useless for the attacker. Crucially, this collapse mechanism remains dormant during benign fine-tuning, ensuring the model's utility and general capabilities are preserved for legitimate users. Extensive empirical results demonstrate that CTRAP effectively counters harmful fine-tuning risks across various LLMs and attack settings, while maintaining high performance in benign scenarios. Our code is available at https://anonymous.4open.science/r/CTRAP.

Biao Yi, Zekun Fei, Jianing Geng et al.

Large reasoning models (LRMs) have emerged as a significant advancement in artificial intelligence, representing a specialized class of large language models (LLMs) designed to tackle complex reasoning tasks. The defining characteristic of LRMs lies in their extensive chain-of-thought (CoT) reasoning capabilities. In this paper, we identify a previously unexplored attack vector against LRMs, which we term "overthinking backdoors". We advance this concept by proposing a novel tunable backdoor, which moves beyond simple on/off attacks to one where an attacker can precisely control the extent of the model's reasoning verbosity. Our attack is implemented through a novel data poisoning methodology. It pairs a tunable trigger-where the number of repetitions signals the desired intensity-with a correspondingly verbose CoT response. These responses are programmatically generated by instructing a teacher LLM to inject a controlled number of redundant refinement steps into a correct reasoning process. The approach preserves output correctness, which ensures stealth and establishes the attack as a pure resource-consumption vector. Extensive empirical results on various LRMs demonstrate that our method can reliably trigger a controllable, multi-fold increase in the length of the reasoning process, without degrading the final answer's correctness. Our source code is available at https://github.com/FZaKK/BadReasoner.

Baolei Zhang, Haoran Xin, Yuxi Chen et al.

Retrieval-Augmented Generation (RAG) integrates external knowledge into large language models to improve response quality. However, recent work has shown that RAG systems are highly vulnerable to poisoning attacks, where malicious texts are inserted into the knowledge database to influence model outputs. While several defenses have been proposed, they are often circumvented by more adaptive or sophisticated attacks. This paper presents RAGOrigin, a black-box responsibility attribution framework designed to identify which texts in the knowledge database are responsible for misleading or incorrect generations. Our method constructs a focused attribution scope tailored to each misgeneration event and assigns a responsibility score to each candidate text by evaluating its retrieval ranking, semantic relevance, and influence on the generated response. The system then isolates poisoned texts using an unsupervised clustering method. We evaluate RAGOrigin across seven datasets and fifteen poisoning attacks, including newly developed adaptive poisoning strategies and multi-attacker scenarios. Our approach outperforms existing baselines in identifying poisoned content and remains robust under dynamic and noisy conditions. These results suggest that RAGOrigin provides a practical and effective solution for tracing the origins of corrupted knowledge in RAG systems. Our code is available at: https://github.com/zhangbl6618/RAG-Responsibility-Attribution

Jianing Geng, Biao Yi, Zekun Fei et al.

Jailbreak attacks pose a serious threat to large language models (LLMs) by bypassing built-in safety mechanisms and leading to harmful outputs. Studying these attacks is crucial for identifying vulnerabilities and improving model security. This paper presents a systematic survey of jailbreak methods from the novel perspective of stealth. We find that existing attacks struggle to simultaneously achieve toxic stealth (concealing toxic content) and linguistic stealth (maintaining linguistic naturalness). Motivated by this, we propose StegoAttack, a fully stealthy jailbreak attack that uses steganography to hide the harmful query within benign, semantically coherent text. The attack then prompts the LLM to extract the hidden query and respond in an encrypted manner. This approach effectively hides malicious intent while preserving naturalness, allowing it to evade both built-in and external safety mechanisms. We evaluate StegoAttack on four safety-aligned LLMs from major providers, benchmarking against eight state-of-the-art methods. StegoAttack achieves an average attack success rate (ASR) of 92.00%, outperforming the strongest baseline by 11.0%. Its ASR drops by less than 1% even under external detection (e.g., Llama Guard). Moreover, it attains the optimal comprehensive scores on stealth detection metrics, demonstrating both high efficacy and exceptional stealth capabilities. The code is available at https://anonymous.4open.science/r/StegoAttack-Jail66

Zekun Fei, Biao Yi, Jianing Geng et al.

Embedding-as-a-Service (EaaS) has emerged as a successful business pattern but faces significant challenges related to various forms of copyright infringement, particularly, the API misuse and model extraction attacks. Various studies have proposed backdoor-based watermarking schemes to protect the copyright of EaaS services. In this paper, we reveal that previous watermarking schemes possess semantic-independent characteristics and propose the Semantic Perturbation Attack (SPA). Our theoretical and experimental analysis demonstrate that this semantic-independent nature makes current watermarking schemes vulnerable to adaptive attacks that exploit semantic perturbations tests to bypass watermark verification. Extensive experimental results across multiple datasets demonstrate that the True Positive Rate (TPR) for identifying watermarked samples under SPA can reach up to more than 95\%, rendering watermarks ineffective while maintaining the high utility of embeddings. Furthermore, we discuss potential defense strategies to mitigate SPA. Our code is available at https://github.com/Zk4-ps/EaaS-Embedding-Watermark.

Baolei Zhang, Yuxi Chen, Minghong Fang et al.

Large language models (LLMs) have demonstrated impressive natural language processing abilities but face challenges such as hallucination and outdated knowledge. Retrieval-Augmented Generation (RAG) has emerged as a state-of-the-art approach to mitigate these issues. While RAG enhances LLM outputs, it remains vulnerable to poisoning attacks. Recent studies show that injecting poisoned text into the knowledge database can compromise RAG systems, but most existing attacks assume that the attacker can insert a sufficient number of poisoned texts per query to outnumber correct-answer texts in retrieval, an assumption that is often unrealistic. To address this limitation, we propose CorruptRAG, a practical poisoning attack against RAG systems in which the attacker injects only a single poisoned text, enhancing both feasibility and stealth. Extensive experiments across multiple datasets demonstrate that CorruptRAG achieves higher attack success rates compared to existing baselines.

Baolei Zhang, Haoran Xin, Jiatong Li et al.

Retrieval-Augmented Generation (RAG) has proven effective in mitigating hallucinations in large language models by incorporating external knowledge during inference. However, this integration introduces new security vulnerabilities, particularly to poisoning attacks. Although prior work has explored various poisoning strategies, a thorough assessment of their practical threat to RAG systems remains missing. To address this gap, we propose the first comprehensive benchmark framework for evaluating poisoning attacks on RAG. Our benchmark covers 5 standard question answering (QA) datasets and 10 expanded variants, along with 13 poisoning attack methods and 7 defense mechanisms, representing a broad spectrum of existing techniques. Using this benchmark, we conduct a comprehensive evaluation of all included attacks and defenses across the full dataset spectrum. Our findings show that while existing attacks perform well on standard QA datasets, their effectiveness drops significantly on the expanded versions. Moreover, our results demonstrate that various advanced RAG architectures, such as sequential, branching, conditional, and loop RAG, as well as multi-turn conversational RAG, multimodal RAG systems, and RAG-based LLM agent systems, remain susceptible to poisoning attacks. Notably, current defense techniques fail to provide robust protection, underscoring the pressing need for more resilient and generalizable defense strategies.

Biao Yi, Jiahao Li, Baolei Zhang et al.

Fine-tuning-as-a-Service introduces a critical vulnerability where a few malicious examples mixed into the user's fine-tuning dataset can compromise the safety alignment of Large Language Models (LLMs). While a recognized paradigm frames safe fine-tuning as a multi-objective optimization problem balancing user task performance with safety alignment, we find existing solutions are critically sensitive to the harmful ratio, with defenses degrading sharply as harmful ratio increases. We diagnose that this failure stems from conflicting gradients, where the user-task update directly undermines the safety objective. To resolve this, we propose SafeGrad, a novel method that employs gradient surgery. When a conflict is detected, SafeGrad nullifies the harmful component of the user-task gradient by projecting it onto the orthogonal plane of the alignment gradient, allowing the model to learn the user's task without sacrificing safety. To further enhance robustness and data efficiency, we employ a KL-divergence alignment loss that learns the rich, distributional safety profile of the well-aligned foundation model. Extensive experiments show that SafeGrad provides state-of-the-art defense across various LLMs and datasets, maintaining robust safety even at high harmful ratios without compromising task fidelity.