Kedong Xiu

Churui Zeng, Weiwei Qi, Kedong Xiu et al.

The rise of LLM agents introduces a new threat by enabling planning, coding, and even end-to-end execution of expert-level attack workflows. However, this threat remains underexplored and underestimated since (i) safety alignment prevents LLMs from directly generating harmful instructions, and (ii) most existing jailbreak methods cannot consistently induce agents to execute malicious operations. In this paper, we propose TRACE, a practical agentic jailbreaking framework to further reveal the risks of this threat surface. To conceal the malicious intent, TRACE decomposes a malicious task into multiple subtask sequences under different schemes and selects the sequence with the fewest explicitly harmful subtasks. TRACE then disguises the remaining harmful subtasks as benign-looking instructions by embedding them in task-aware scenarios with related roles, environments, directives, and heuristics. The scenarios are iteratively evolved through well-defined transformation actions, which are sampled by a Q-learning-inspired mechanism, for inducing the agent to execute on the harmful subtasks. Extensive evaluations on AgentHarm and AdvCUA show that TRACE consistently outperforms existing jailbreak baselines across multiple advanced LLM agents, achieving up to 100% bypass rate and 0.73 average success score. We also demonstrate the effectiveness of TRACE in controlled cyberattack instances. Our code and demos are available at https://github.com/ZJU-LLM-Safety/TRACE.git.

Bo Lv, Zhiheng Xu, KeDong Xiu et al.

Mixture-of-Experts (MoE) architectures have become an increasingly important paradigm for scaling Large Language Models (LLMs). As MoE models are increasingly deployed in real-world services, safety auditing becomes necessary to verify whether these models produce or facilitate harmful behaviors during operation. However, existing content-based auditing methods typically require access to user prompts, model inputs, or generated outputs, potentially exposing sensitive user information and creating a fundamental tension between LLM safety and user privacy. On the other hand, we observe that, in MoE models, sparse expert routing maps different inputs to activate different expert-execution patterns, producing measurable footprints in low-level GPU execution telemetry. Inspired by this observation, we propose RouteScan, a non-intrusive auditing framework for detecting harmful behaviors through GPU-level expert routing telemetry. Specifically, RouteScan utilizes the number of active GPU threads allocated to expert modules during the prefilling phase as a discriminative micro-architectural fingerprint, and builds a lightweight detection pipeline that isolates cross-domain invariant risk indicators for the precise identification of malicious prompts. Comprehensive evaluations on open-source MoE LLMs with distinct routing designs demonstrate that RouteScan achieves strong generalization, with an AUROC exceeding 0.93 on unseen harmful domains and 0.96 under novel jailbreak wrappers. Moreover, empirical inversion tests show that the collected expert routing telemetry provides limited information for prompt reconstruction, suggesting a practical privacy advantage over content-based auditing methods.

Xinzhe Huang, Kedong Xiu, Tianhang Zheng et al.

Recent research has focused on exploring the vulnerabilities of Large Language Models (LLMs), aiming to elicit harmful and/or sensitive content from LLMs. However, due to the insufficient research on dual-jailbreaking -- attacks targeting both LLMs and Guardrails, the effectiveness of existing attacks is limited when attempting to bypass safety-aligned LLMs shielded by guardrails. Therefore, in this paper, we propose DualBreach, a target-driven framework for dual-jailbreaking. DualBreach employs a Target-driven Initialization (TDI) strategy to dynamically construct initial prompts, combined with a Multi-Target Optimization (MTO) method that utilizes approximate gradients to jointly adapt the prompts across guardrails and LLMs, which can simultaneously save the number of queries and achieve a high dual-jailbreaking success rate. For black-box guardrails, DualBreach either employs a powerful open-sourced guardrail or imitates the target black-box guardrail by training a proxy model, to incorporate guardrails into the MTO process. We demonstrate the effectiveness of DualBreach in dual-jailbreaking scenarios through extensive evaluation on several widely-used datasets. Experimental results indicate that DualBreach outperforms state-of-the-art methods with fewer queries, achieving significantly higher success rates across all settings. More specifically, DualBreach achieves an average dual-jailbreaking success rate of 93.67% against GPT-4 with Llama-Guard-3 protection, whereas the best success rate achieved by other methods is 88.33%. Moreover, DualBreach only uses an average of 1.77 queries per successful dual-jailbreak, outperforming other state-of-the-art methods. For the purpose of defense, we propose an XGBoost-based ensemble defensive mechanism named EGuard, which integrates the strengths of multiple guardrails, demonstrating superior performance compared with Llama-Guard-3.

Kedong Xiu, Sai Qian Zhang

As Vision-Language Models (VLMs) are increasingly deployed in split-DNN configurations--with visual encoders (e.g., ResNet, ViT) operating on user devices and sending intermediate features to the cloud--there is a growing privacy risk from semantic information leakage. Existing approaches to reconstructing images from these intermediate features often result in blurry, semantically ambiguous images. To directly address semantic leakage, we propose CapRecover, a cross-modality inversion framework that recovers high-level semantic content, such as labels or captions, directly from intermediate features without image reconstruction. We evaluate CapRecover on multiple datasets and victim models, demonstrating strong performance in semantic recovery. Specifically, CapRecover achieves up to 92.71% Top-1 label accuracy on CIFAR-10 and generates fluent captions from ResNet50 features on COCO2017 with ROUGE-L scores up to 0.52. Our analysis further reveals that deeper convolutional layers encode significantly more semantic information compared to shallow layers. To mitigate semantic leakage, we introduce a simple yet effective protection method: adding random noise to intermediate features at each layer and removing the noise in the next layer. Experimental results show that this approach prevents semantic leakage without additional training costs. Our code is available at https://jus1mple.github.io/Image2CaptionAttack.

Langqi Yang, Tianhang Zheng, Kedong Xiu et al.

The alignment of large language models (LLMs) with human values is critical for their safe deployment, yet jailbreak attacks can subvert this alignment to elicit harmful outputs from LLMs. In recent years, a proliferation of jailbreak attacks has emerged, accompanied by diverse metrics and judges to assess the harmfulness of the LLM outputs. However, the absence of a systematic benchmark to assess the quality and effectiveness of these metrics and judges undermines the credibility of the reported jailbreak effectiveness and other risks. To address this gap, we introduce HarmMetric Eval, a comprehensive benchmark designed to support both overall and fine-grained evaluation of harmfulness metrics and judges. Our benchmark includes a high-quality dataset of representative harmful prompts paired with diverse harmful and non-harmful model responses, alongside a flexible scoring mechanism compatible with various metrics and judges. With HarmMetric Eval, our extensive experiments uncover a surprising result: two conventional metrics--METEOR and ROUGE-1--outperform LLM-based judges in evaluating the harmfulness of model responses, challenging prevailing beliefs about LLMs' superiority in this domain. Our dataset is publicly available at https://huggingface.co/datasets/qusgo/HarmMetric_Eval, and the code is available at https://anonymous.4open.science/r/HarmMetric-Eval-4CBE.

Xinzhe Huang, Wenjing Hu, Tianhang Zheng et al.

Existing gradient-based jailbreak attacks on Large Language Models (LLMs), such as Greedy Coordinate Gradient (GCG) and COLD-Attack, typically optimize adversarial suffixes to align the LLM output with a predefined target response. However, by restricting the optimization objective as inducing a predefined target, these methods inherently constrain the adversarial search space, which limit their overall attack efficacy. Furthermore, existing methods typically require a large number of optimization iterations to fulfill the large gap between the fixed target and the original model response, resulting in low attack efficiency. To overcome the limitations of targeted jailbreak attacks, we propose the first gradient-based untargeted jailbreak attack (UJA), aiming to elicit an unsafe response without enforcing any predefined patterns. Specifically, we formulate an untargeted attack objective to maximize the unsafety probability of the LLM response, which can be quantified using a judge model. Since the objective is non-differentiable, we further decompose it into two differentiable sub-objectives for optimizing an optimal harmful response and the corresponding adversarial prompt, with a theoretical analysis to validate the decomposition. In contrast to targeted jailbreak attacks, UJA's unrestricted objective significantly expands the search space, enabling a more flexible and efficient exploration of LLM vulnerabilities.Extensive evaluations demonstrate that UJA can achieve over 80% attack success rates against recent safety-aligned LLMs with only 100 optimization iterations, outperforming the state-of-the-art gradient-based attacks such as I-GCG and COLD-Attack by over 20%.

Kedong Xiu, Churui Zeng, Tianhang Zheng et al.

Existing gradient-based jailbreak attacks typically optimize an adversarial suffix to induce a fixed affirmative response. However, this fixed target usually resides in an extremely low-density region of a safety-aligned LLM's output distribution conditioned on diverse harmful inputs. Due to the substantial discrepancy between the target and the original output, existing attacks require numerous iterations to optimize the adversarial prompt, which might still fail to induce the low-probability target response from the target LLM. In this paper, we propose Dynamic Target Attack (DTA), a new jailbreaking framework relying on the target LLM's own responses as targets to optimize the adversarial prompts. In each optimization round, DTA iteratively samples multiple candidate responses directly from the output distribution conditioned on the current prompt, and selects the most harmful response as a temporary target for prompt optimization. In contrast to existing attacks, DTA significantly reduces the discrepancy between the target and the output distribution, substantially easing the optimization process to search for an effective adversarial prompt. Extensive experiments demonstrate the superior effectiveness and efficiency of DTA: under the white-box setting, DTA only needs 200 optimization iterations to achieve an average attack success rate (ASR) of over 87\% on recent safety-aligned LLMs, exceeding the state-of-the-art baselines by over 15\%. The time cost of DTA is 2-26 times less than existing baselines. Under the black-box setting, DTA uses Llama-3-8B-Instruct as a surrogate model for target sampling and achieves an ASR of 85\% against the black-box target model Llama-3-70B-Instruct, exceeding its counterparts by over 25\%.