arXiv subjects

Xiangtao Meng

Publications and source records attributed to Xiangtao Meng.

14 recordsLinked to original sources

arXiv2026

Defenses at Odds: Measuring and Explaining Defense Conflicts in Large Language Models

Large language models (LLMs) may require additional defenses after deployment as risks and governance requirements evolve. Subsequent defenses can interact with earlier defenses, raising the question of their sequential compatibility. We study this question with CONFLICTEVAL, evaluating 144 ordered compositions of six defenses spanning safety, privacy, and fairness across six models. The resulting interactions are heterogeneous across defense pairs, application orders, and models. Notably, some compositions exhibit defense conflicts: the subsequent defense improves its target objective while weakening protection established by the earlier defense. We investigate these interactions through the directional compatibility of defense-induced changes in risk-relevant representations. Across 11 selected cases, we use activation interventions to assess which defense-induced representational changes support protection and examine how subsequent defenses affect these changes. We find that, in some conflict cases, subsequent defenses counteract representational changes supporting earlier protection, providing evidence for one possible pathway to defense conflicts. Building on this analysis, we propose Conflict-Triggered Directional Retention (CTDR), which penalizes opposing shifts along directions supporting earlier protection. On six selected conflicting compositions from this analysis, CTDR reduces first-objective regression while maintaining positive gains on the subsequent defense objective.

cs.CR↗

arXiv2026

SceneJail: Exploiting Video Scenario Context to Jailbreak Multimodal LLMs

Video Multimodal Large Language Models (Video-MLLMs) support reasoning over video inputs, yet remain vulnerable to jailbreak attacks that elicit policy-violating responses. Existing video jailbreaks primarily manipulate how harmful queries are visually presented, thereby treating video merely as a carrier. Consequently, the surrounding video scenario remains unexplored as a contextual attack surface. In this paper, we show that the same harmful query can elicit different safety responses when placed in different video scenarios. To systematically exploit this vulnerability, we propose SceneJail, an adaptive black-box jailbreak framework with two coordinated components. Adaptive Scenario Construction dynamically searches for a surrounding scenario that is contextually compatible with the harmful query. Scenario-aware Prompt Search uses black-box response feedback to search for textual guidance tailored to the selected scenario. Extensive evaluations on the HADES and SafeBench datasets across eight Video-MLLMs, including two proprietary models, GPT-4.1 and Gemini3.5-Flash, demonstrate the effectiveness of SceneJail. SceneJail-F, which presents the complete query persistently, achieves average attack success rates (ASR) up to 91.5%, outperforming the strongest baselines by 29.1 percentage points. Furthermore, SceneJail-S, which distributes the query across successive frames, remains highly robust against current defenses, retaining a 72.3% ASR even under strict image filtering.

cs.CR↗

arXiv2026

One Risk Down, Another Up: Cross-Risk Interactions Induced by LLM Defenses

Large Language Models (LLMs) are increasingly deployed in high-stakes settings, where they face diverse risks. Numerous defense strategies have been proposed to mitigate these risks, but they are almost always evaluated in isolation. This isolated view leaves a critical question open: does mitigating one risk inadvertently change a model's exposure to others? Beyond the well-studied risk-utility trade-off, we present the first systematic study of cross-risk interactions induced by LLM defenses. We propose CrossRiskEval, an evaluation paradigm that situates a defended model in a multi-dimensional risk space and quantifies how a defense built for one risk shifts the others. Among 166 cross-risk evaluations covering 32 defended models, 77.1% exhibit statistically significant cross-risk interactions. Most of these interactions amplify non-target risks, with increases exceeding 100% in some cases. Beyond behavioral evaluation, we conduct neuron-level analyses in seven selected cases to investigate one possible pathway associated with these interactions. We identify conflict-entangled neurons whose activation interventions produce opposing effects on proxies for the target and non-target risks. In conflict cases, restoring these neurons to their base-model activations partially reduces the corresponding risk increases, providing evidence that defense-induced changes to these neurons may contribute to the observed interactions. Building on this evidence, we propose Conflict-Aware Freezing, a training-time strategy that prevents direct updates to the parameters associated with the identified neurons. Across five conflict cases, it offsets 35%-196% of non-target risk amplification while meeting the original defense criterion.

cs.CR↗

arXiv2026

Transferable End-to-End Optimization for Indirect Long-Term Memory Poisoning in LLM Agents

Long-term memory can turn untrusted external content into persistent influence over an LLM agent's future decisions, creating the threat of indirect memory poisoning. A successful attack must survive a multi-stage pipeline comprising memory writing, retrieval, and utilization. Existing attacks largely rely on intra-stage optimization, optimizing individual stages in isolation while overlooking inter-stage coupling. Specifically, these stages impose different requirements on the same poisoning content, and each stage operates on the transformed output of its predecessor. Consequently, optimizing one stage may undermine the effectiveness of other stages, while upstream transformations may erase improvements intended for downstream stages. Indirect memory poisoning should therefore be viewed as an end-to-end optimization problem. Based on this insight, we present \textsc{PipePoison}, which collects fine-grained stage feedback from local shadow systems, uses chain-structured losses to identify and optimize the stage bottlenecking end-to-end success, and applies stability-calibrated stage and configuration weights to improve transferability. Across three agent frameworks and four memory mechanisms, \textsc{PipePoison} improves attack utilization rate by 19.1 percentage points. Even on fully unseen victim configurations, it outperforms the strongest baseline by 16 percentage points and remains effective under eight representative defenses.

cs.CR↗

arXiv2026

Understanding Stage-Wise Utility-Risk Trade-offs in LLM Agent Memory

Long-term memory is becoming a core capability of LLM agents, enabling personalization and long-horizon interaction. However, memory mechanisms that retain, transform, or expose more information can affect both benign utility and susceptibility to memory poisoning. Existing evaluations typically measure memory utility or attack risk in isolation under fixed configurations, providing limited insight into how stage-specific design choices reshape their trade-off. We present \textsc{MemGauge}, a controllable framework that separately varies writing admission, management policy, and retrieval exposure under matched clean and poisoned conditions. Across 11 LLMs and two long-term memory benchmarks, controlled evaluations reveal three distinct profiles: a threshold-like risk transition during writing, policy-dependent local decoupling during management, and coupled growth of utility and risk during retrieval. We further apply analogous stage-level measurements to four existing memory systems and observe diagnostic associations qualitatively consistent with these profiles. These results show that targeted poisoning risk varies across memory operations and motivate stage-aware evaluation and control of LLM-agent memory.

cs.CR↗

arXiv2026

Extracting Knowledge from Tools in LLM Agents

LLM agents commonly use knowledge-based tools and access their underlying files, databases, and search indexes through tool invocation. This integration improves agents' ability to provide domain-specific services but also introduces the risk of tool-mediated knowledge extraction: source content exposed to an agent for legitimate responses may be progressively recovered from its outputs, enabling reconstruction of the knowledge source behind a target tool. This paper systematically investigates this risk and identifies two challenges introduced by tool invocation: tool-selection uncertainty, where an agent may invoke a competing tool instead of the target tool, and tool-argument compression, where fine-grained query information may be lost when the agent generates tool arguments. To tackle these challenges, we propose ToolSiphon, a query-only extraction attack that introduces two complementary signals: a target-discriminative signal, implemented through Tool Contrastive Analysis, to steer queries toward the target tool; and a response-grounded factual signal, implemented through Evidence Chained Feedback, to mitigate argument compression and progressively expand extraction coverage. Across three types of knowledge-based tools and six domain-specific datasets, ToolSiphon recovers 74.3% of source records on average when coarse-grained information about non-target tools is available, with 83.2% textual recovery and 90.2% semantic similarity. Even without such information, it recovers 66.3% of source records. ToolSiphon also remains effective against representative defenses and on three real-world agent platforms.

cs.CR↗

arXiv2026

Understanding and Exploiting Initialization Anchoring Weakness in Feedback-Based Agent Planning

Feedback-based planning improves agent reliability by incorporating tool observations and corrective feedback. However, its protection may not be distributed uniformly across planning stages. We conduct a round-wise analysis of four representative feedback mechanisms and uncover an initialization anchoring weakness: the first feedback round corrects 46\% of adversarial directions, whereas the rates fall to 13\% and 7\% among directions surviving into the next two rounds. Our analysis attributes this weakness to three interacting factors: a contextually plausible shift in the initial plan, insufficient counterevidence, and the persistence of accepted directions in the accumulated trajectory. Based on these findings, we propose \textsc{InitAnchor}, a black-box framework for exploiting this weakness through attacker-controlled external materials. It operationalizes the three factors as directional-shift, contextual-plausibility, and counterevidence-resilience signals under either limited target access or no target access. Across 112 tasks from 16 domains, six agent architectures, and five backbone LLMs, \textsc{InitAnchor} achieves average ASRs of 76.1\% and 72.0\% under the two settings while reducing first-round mitigation rates to 21.0\% and 25.0\%, respectively. It also remains effective against six defenses and across six real-world agent systems. These findings show that feedback-based agents can retain early biases even when later correction is available.

cs.CR↗

arXiv2026

Isolated but Exposed: Persistence-Based Memory Extraction Attack on LLM Agents

LLM-based agents extend large language models with long-term memory (LTM) that persists privacy-sensitive user data across sessions. Production systems mitigate extraction risks through memory isolation, binding each user's LTM to a unique identifier. This defense has blocked known attacks on shared storage, fostering the assumption that isolated LTM is secure. We identify the tool interface as an overlooked attack surface. Agents routinely embed LTM-retrieved data in tool invocation parameters, enabling a malicious tool to exfiltrate private memory without violating user-level isolation. Naive adaptations of user-side extraction techniques fail because the adversarial command's semantics interfere with retrieval precision, and platform-imposed tool-call limits constrain the extraction budget per trigger. We present SPORE, the first extraction attack designed for this threat model. SPORE decouples the adversarial command from retrieval anchors by persisting the command in short-term memory and emitting semantically pure anchors in tool responses. The restored retrieval precision enables a geometric coverage optimization over the embedding space that systematically steers anchors toward unexplored memory regions. To sustain extraction beyond tool-call limits, SPORE persists reactivation payloads in memory that automatically resume the attack within and across sessions without additional user triggers. SPORE achieves an 80.0% record extraction rate with unlimited triggers and 47.0% with only 20 triggers. In multi-user deployments, attackers can link extracted records to user identities, enabling targeted surveillance. These results demonstrate that memory isolation alone is insufficient and call for reexamining tool-side trust boundaries in agent architectures.

cs.CR↗

arXiv2026

Not All Tokens Are Created Equal: Query-Efficient Jailbreak Fuzzing for LLMs

Large Language Models(LLMs) are widely deployed, yet are vulnerable to jailbreak prompts that elicit policy-violating outputs. Although prior studies have uncovered these risks, they typically treat all tokens as equally important during prompt mutation, overlooking the varying contributions of individual tokens to triggering model refusals. Consequently, these attacks introduce substantial redundant searching under query-constrained scenarios, reducing attack efficiency and hindering comprehensive vulnerability assessment. In this work, we conduct a token-level analysis of refusal behavior and observe that token contributions are highly skewed rather than uniform. Moreover, we find strong cross-model consistency in refusal tendencies, enabling the use of a surrogate model to estimate token-level contributions to the target model's refusals. Motivated by these findings, we propose TriageFuzz, a token-aware jailbreak fuzzing framework that adapts the fuzz testing approach with a series of customized designs. TriageFuzz leverages a surrogate model to estimate the contribution of individual tokens to refusal behaviors, enabling the identification of sensitive regions within the prompt. Furthermore, it incorporates a refusal-guided evolutionary strategy that adaptively weights candidate prompts with a lightweight scorer to steer the evolution toward bypassing safety constraints. Extensive experiments on six open-source LLMs and three commercial APIs demonstrate that TriageFuzz achieves comparable attack success rates (ASR) with significantly reduced query costs. Notably, it attains a 90% ASR with over 70% fewer queries compared to baselines. Even under an extremely restrictive budget of 25 queries, TriageFuzz outperforms existing methods, improving ASR by 20-40%.

cs.CR↗

arXiv2026

Beyond the Safety Tax: Mitigating Unsafe Text-to-Image Generation via External Safety Rectification

Text-to-image (T2I) generative models have achieved remarkable visual fidelity, yet remain vulnerable to generating unsafe content. Existing safety defenses typically intervene internally within the generative model, but suffer from severe concept entanglement, leading to degradation of benign generation quality, a trade-off we term the Safety Tax. To overcome this limitation, we advocate a paradigm shift from destructive internal editing to external safety rectification. Following this principle, we propose SafePatch, a structurally isolated safety module that performs external, interpretable rectification without modifying the base model. The core backbone of SafePatch is architecturally instantiated as a trainable clone of the base model's encoder, allowing it to inherit rich semantic priors and maintain representation consistency. To enable interpretable safety rectification, we construct a strictly aligned counterfactual safety dataset (ACS) for differential supervision training. Across nudity and multi-category benchmarks and recent adversarial prompt attacks, SafePatch achieves robust unsafe suppression (7% unsafe on I2P) while preserving image quality and semantic alignment.

cs.CV↗

arXiv2026

Beyond Known Fakes: Generalized Detection of AI-Generated Images via Post-hoc Distribution Alignment

The rapid proliferation of highly realistic AI-generated images poses serious security threats such as misinformation and identity fraud. Detecting generated images in open-world settings is particularly challenging when they originate from unknown generators, as existing methods typically rely on model-specific artifacts and require retraining on new fake data, limiting their generalization and scalability. In this work, we propose Post-hoc Distribution Alignment (PDA), a generalized and model-agnostic framework for detecting AI-generated images under unknown generative threats. Specifically, PDA reformulates detection as a distribution alignment task by regenerating test images through a known generative model. When real images are regenerated, they inherit model-specific artifacts and align with the known fake distribution. In contrast, regenerated unknown fakes contain incompatible or mixed artifacts and remain misaligned. This difference allows an existing detector, trained on the known generative model, to accurately distinguish real images from unknown fakes without requiring access to unseen data or retraining. Extensive experiments across 16 state-of-the-art generative models, including GANs, diffusion models, and commercial text-to-image APIs (e.g., Midjourney), demonstrate that PDA achieves average detection accuracy of 96.69%, outperforming the best baseline by 10.71%. Comprehensive ablation studies and robustness analyses further confirm PDA's generalizability and resilience to distribution shifts and image transformations. Overall, our work provides a practical and scalable solution for real-world AI-generated image detection where new generative models emerge continuously.

cs.CR↗

arXiv2025

DCMI: A Differential Calibration Membership Inference Attack Against Retrieval-Augmented Generation

While Retrieval-Augmented Generation (RAG) effectively reduces hallucinations by integrating external knowledge bases, it introduces vulnerabilities to membership inference attacks (MIAs), particularly in systems handling sensitive data. Existing MIAs targeting RAG's external databases often rely on model responses but ignore the interference of non-member-retrieved documents on RAG outputs, limiting their effectiveness. To address this, we propose DCMI, a differential calibration MIA that mitigates the negative impact of non-member-retrieved documents. Specifically, DCMI leverages the sensitivity gap between member and non-member retrieved documents under query perturbation. It generates perturbed queries for calibration to isolate the contribution of member-retrieved documents while minimizing the interference from non-member-retrieved documents. Experiments under progressively relaxed assumptions show that DCMI consistently outperforms baselines--for example, achieving 97.42% AUC and 94.35% Accuracy against the RAG system with Flan-T5, exceeding the MBA baseline by over 40%. Furthermore, on real-world RAG platforms such as Dify and MaxKB, DCMI maintains a 10%-20% advantage over the baseline. These results highlight significant privacy risks in RAG systems and emphasize the need for stronger protection mechanisms. We appeal to the community's consideration of deeper investigations, like ours, against the data leakage risks in rapidly evolving RAG systems. Our code is available at https://github.com/Xinyu140203/RAG_MIA.

cs.CR↗

arXiv2025

Fuzz-Testing Meets LLM-Based Agents: An Automated and Efficient Framework for Jailbreaking Text-To-Image Generation Models

Text-to-image (T2I) generative models have revolutionized content creation by transforming textual descriptions into high-quality images. However, these models are vulnerable to jailbreaking attacks, where carefully crafted prompts bypass safety mechanisms to produce unsafe content. While researchers have developed various jailbreak attacks to expose this risk, these methods face significant limitations, including impractical access requirements, easily detectable unnatural prompts, restricted search spaces, and high query demands on the target system. In this paper, we propose JailFuzzer, a novel fuzzing framework driven by large language model (LLM) agents, designed to efficiently generate natural and semantically meaningful jailbreak prompts in a black-box setting. Specifically, JailFuzzer employs fuzz-testing principles with three components: a seed pool for initial and jailbreak prompts, a guided mutation engine for generating meaningful variations, and an oracle function to evaluate jailbreak success. Furthermore, we construct the guided mutation engine and oracle function by LLM-based agents, which further ensures efficiency and adaptability in black-box settings. Extensive experiments demonstrate that JailFuzzer has significant advantages in jailbreaking T2I models. It generates natural and semantically coherent prompts, reducing the likelihood of detection by traditional defenses. Additionally, it achieves a high success rate in jailbreak attacks with minimal query overhead, outperforming existing methods across all key metrics. This study underscores the need for stronger safety mechanisms in generative models and provides a foundation for future research on defending against sophisticated jailbreaking attacks. JailFuzzer is open-source and available at this repository: https://github.com/YingkaiD/JailFuzzer.

cs.CR↗

arXiv2023

AVA: Inconspicuous Attribute Variation-based Adversarial Attack bypassing DeepFake Detection

While DeepFake applications are becoming popular in recent years, their abuses pose a serious privacy threat. Unfortunately, most related detection algorithms to mitigate the abuse issues are inherently vulnerable to adversarial attacks because they are built atop DNN-based classification models, and the literature has demonstrated that they could be bypassed by introducing pixel-level perturbations. Though corresponding mitigation has been proposed, we have identified a new attribute-variation-based adversarial attack (AVA) that perturbs the latent space via a combination of Gaussian prior and semantic discriminator to bypass such mitigation. It perturbs the semantics in the attribute space of DeepFake images, which are inconspicuous to human beings (e.g., mouth open) but can result in substantial differences in DeepFake detection. We evaluate our proposed AVA attack on nine state-of-the-art DeepFake detection algorithms and applications. The empirical results demonstrate that AVA attack defeats the state-of-the-art black box attacks against DeepFake detectors and achieves more than a 95% success rate on two commercial DeepFake detectors. Moreover, our human study indicates that AVA-generated DeepFake images are often imperceptible to humans, which presents huge security and privacy concerns.

cs.CV↗