Autonomous agent frameworks built upon large language models (LLMs) are evolving into complex, tool-integrated, and continuously operating systems, introducing security risks beyond traditional prompt-level vulnerabilities. As this paradigm is still at an early stage of development, a timely and systematic understanding of its security implications is increasingly important. Although a growing body of work has examined different attack surfaces and defense problems in agent systems, existing studies remain scattered across individual aspects of agent security, and there is still a lack of a layered review on this topic. To address this gap, this survey presents a layered review of security risks and defense strategies in autonomous agent frameworks, with OpenClaw as a case study. We organize the analysis into four security-relevant layers: the context and instruction layer, the tool and action layer, the state and persistence layer, and the ecosystem and automation layer. For each layer, we summarize its functional role, representative security risks, and corresponding defense strategies. Based on this layered analysis, we further identify that threats in autonomous agent frameworks may propagate across layers, from manipulated inputs to unsafe actions, persistent state contamination, and broader ecosystem-level impact. Finally, we highlight potential key challenges, including research imbalance across layers, the lack of long-horizon evaluation, and weak ecosystem trust models, and outline future directions toward more systematic and integrated defenses.

Constructing and curating high-quality code datasets requires significant resources, making them valuable intellectual property. Unfortunately, these datasets currently face severe risks of unauthorized use. Although digital watermarking offers a post hoc mechanism for copyright authentication, existing methods are predominantly based on the co-occurrence pattern, which is not robust and is susceptible to watermark detection and removal attacks. In this paper, we propose PuzzleMark, a robust watermarking method for code datasets. To reduce the risk of watermark exposure, PuzzleMark introduces a carrier selection strategy that leverages code complexity to evaluate the suitability of code snippets as watermark carriers, and selects those with high suitability for watermarking. To enhance the robustness of the watermark, PuzzleMark proposes a novel concatenation pattern to replace the traditional co-occurrence pattern, and implements two watermarking strategies through variable name concatenation. PuzzleMark adaptively embeds watermarks based on the inherent characteristics of the code, making it more stealthy while maintaining design simplicity. For watermark verification, PuzzleMark employs Fisher's exact test to verify suspicious models under a black-box setting. Experimental results demonstrate that PuzzleMark achieves a 100% verification success rate and a 0% false positive rate, with negligible impact on model performance. Both our human study and our evaluation using four state-of-the-art watermark detection methods show that PuzzleMark exhibits strong imperceptibility, with an average suspicious rate $\leq$ 0.24 and an average recall $\leq$ 30.41%, respectively. As a practical digital watermarking method, PuzzleMark provides strong protection for the intellectual property of code datasets and offers new insights for future research.

Local fine-tuning datasets routinely contain sensitive secrets such as API keys, personal identifiers, and financial records. Although ''local offline fine-tuning'' is often viewed as a privacy boundary, we reveal that compromised model code is sufficient to steal them. Current passive pretrained-weight poisoning attacks, while effective for natural language, fundamentally fail to capture such sparse high-entropy targets due to their reliance on probabilistic semantic prefixes. To bridge this gap, we identify and exploit a practical but overlooked supply-chain vector -- model code camouflaged as standard architectural definitions -- to realize a paradigm shift from passive weight poisoning to active execution hijacking. We introduce a deterministic full-chain memorization mechanism: it locks onto token-level secrets in dynamic computation flows via online tensor-rule matching, and leverages value-gradient decoupling to stealthily inject attack gradients, overcoming gradient drowning to force model memorization. Furthermore, we achieve, for the first time, attacker-verifiable secret stealing through black-box queries that precisely distinguishes true leakage from hallucination. Experiments demonstrate that our method achieves over 98\% Strict ASR without compromising the primary task, and can effectively bypass defense measures including DP-SGD, semantic auditing, and code auditing.

The widespread availability of large-scale code datasets has accelerated the development of code large language models (CodeLLMs), raising concerns about unauthorized dataset usage. Dataset poisoning offers a proactive defense by reducing the utility of such unauthorized training. However, existing poisoning methods often require full dataset poisoning and introduce transformations that break code compilability. In this paper, we introduce FunPoison, a functionality-preserving poisoning approach that injects short, compilable weak-use fragments into executed code paths. FunPoison leverages reusable statement-level templates with automatic repair and conservative safety checking to ensure side-effect freedom, while a type-aware synthesis module suppresses static analysis warnings and enhances stealth. Extensive experiments show that FunPoison achieves effective poisoning by contaminating only 10% of the dataset, while maintaining 100% compilability and functional correctness, and remains robust against various advanced code sanitization techniques.

The proliferation of large language models for code (CodeLMs) and open-source contributions has heightened concerns over unauthorized use of source code datasets. While watermarking provides a viable protection mechanism by embedding ownership signals, existing methods rely on detectable trigger-target patterns and are limited to source-code tasks, overlooking other scenarios such as decompilation tasks. In this paper, we propose DuCodeMark, a stealthy and robust dual-purpose watermarking method for code datasets that generalizes across both source-code tasks and decompilation tasks. DuCodeMark parses each code sample into an abstract syntax tree (AST), applies language-specific style transformations to construct stealthy trigger-target pairs, and injects repressible poisoned features into a subset of return-typed samples to enhance robustness against watermark removal or evasion. These features remain inactive during normal training but are activated upon watermark removal, degrading model performance. For verification, DuCodeMark employs a black-box method based on the independent-samples $t$-test. We conduct a comprehensive evaluation of DuCodeMark across 72 settings spanning two code tasks, two programming languages, three CodeLMs, and six decoding temperatures. The results demonstrate that it consistently achieves strong verifiability ($p < 0.05$), high stealthiness (suspicious rate $\leq$ 0.36), robustness against both watermark and poisoning attacks (recall $\leq$ 0.57), and a substantial drop in model performance upon watermark removal (Pass@1 drops by 28.6%), underscoring its practicality and resilience.

DNNs are susceptible to defects like backdoors, adversarial attacks, and unfairness, undermining their reliability. Existing approaches mainly involve retraining, optimization, constraint-solving, or search algorithms. However, most methods rely on gradient calculations, restricting applicability to specific activation functions (e.g., ReLU), or use search algorithms with uninterpretable localization and repair. Furthermore, they often lack generalizability across multiple properties. We propose SHARPEN, integrating interpretable fault localization with a derivative-free optimization strategy. First, SHARPEN introduces a Deep SHAP-based localization strategy quantifying each layer's and neuron's marginal contribution to erroneous outputs. Specifically, a hierarchical coarse-to-fine approach reranks layers by aggregated impact, then locates faulty neurons/filters by analyzing activation divergences between property-violating and benign states. Subsequently, SHARPEN incorporates CMA-ES to repair identified neurons. CMA-ES leverages a covariance matrix to capture variable dependencies, enabling gradient-free search and coordinated adjustments across coupled neurons. By combining interpretable localization with evolutionary optimization, SHARPEN enables derivative-free repair across architectures, being less sensitive to gradient anomalies and hyperparameters. We demonstrate SHARPEN's effectiveness on three repair tasks. Balancing property repair and accuracy preservation, it outperforms baselines in backdoor removal (+10.56%), adversarial mitigation (+5.78%), and unfairness repair (+11.82%). Notably, SHARPEN handles diverse tasks, and its modular design is plug-and-play with different derivative-free optimizers, highlighting its flexibility.

With the development of deep learning, Neural Code Models (NCMs) such as CodeBERT and CodeLlama are widely used for code understanding tasks, including defect detection and code classification. However, recent studies have revealed that NCMs are vulnerable to adversarial examples, inputs with subtle perturbations that induce incorrect predictions while remaining difficult to detect. Existing defenses address this issue via data augmentation to empirically improve robustness, but they are costly, offer no theoretical robustness guarantees, and typically require white-box access to model internals, such as gradients. To address the above challenges, we propose ENBECOME, a novel black-box training-free and lightweight adversarial defense. ENBECOME is designed to both enhance empirical robustness and report certified robustness boundaries for NCMs. ENBECOME operates solely during inference, introducing random, semantics-preserving perturbations to input code snippets to smooth the NCM's decision boundaries. This smoothing enables ENBECOME to formally certify a robustness radius within which adversarial examples can never induce misclassification, a property known as certified robustness. We conduct comprehensive experiments across multiple NCM architectures and tasks. Results show that ENBECOME significantly reduces attack success rates while maintaining high accuracy. For example, in defect detection, it reduces the average ASR from 42.43% to 9.74% with only a 0.29% drop in accuracy. Results show that ENBECOME significantly reduces attack success rates while maintaining high accuracy. For example, in defect detection, it reduces the average ASR from 42.43% to 9.74% with only a 0.29% drop in accuracy. Furthermore, ENBECOME achieves an average certified robustness radius of 1.63, meaning that adversarial modifications to no more than 1.63 identifiers are provably ineffective.

The rapid development of large language models (LLMs) has driven the widespread adoption of cloud-based LLM inference services, while also bringing prominent privacy risks associated with the transmission and processing of private data in remote inference. For privacy-preserving LLM inference technologies to be practically applied in industrial scenarios, three core requirements must be satisfied simultaneously: (1) Accuracy and efficiency losses should be minimized to mitigate degradation in service experience. (2) The inference process can be run on large-scale clusters consist of heterogeneous legacy xPUs. (3) Compatibility with existing LLM infrastructures should be ensured to reuse their engineering optimizations. To the best of our knowledge, none of the existing privacy-preserving LLM inference methods satisfy all the above constraints while delivering meaningful privacy guarantees. In this paper, we propose AloePri, the first privacy-preserving LLM inference method for industrial applications. AloePri protects both the input and output data by covariant obfuscation, which jointly transforms data and model parameters to achieve better accuracy and privacy. We carefully design the transformation for each model component to ensure inference accuracy and data privacy while keeping full compatibility with existing infrastructures of Language Model as a Service. AloePri has been integrated into an industrial system for the evaluation of mainstream LLMs. The evaluation on Deepseek-V3.1-Terminus model (671B parameters) demonstrates that AloePri causes accuracy loss of 0.0%~3.5% and exhibits efficiency equivalent to that of plaintext inference. Meanwhile, AloePri successfully resists state-of-the-art attacks, with less than 5\% of tokens recovered. To the best of our knowledge, AloePri is the first method to exhibit practical applicability to large-scale models in real-world systems.

Federated Learning (FL) enables collaborative model training without exposing clients' private data, and has been widely adopted in privacy-sensitive scenarios. However, FL faces two critical security threats: curious servers that may launch inference attacks to reconstruct clients' private data, and compromised clients that can launch poisoning attacks to disrupt model aggregation. Existing solutions mitigate these attacks by combining mainstream privacy-preserving techniques with defensive aggregation strategies. However, they either incur high computation and communication overhead or perform poorly under non-independent and identically distributed (Non-IID) data settings. To tackle these challenges, we propose SRFed, an efficient Byzantine-robust and privacy-preserving FL framework for Non-IID scenarios. First, we design a decentralized efficient functional encryption (DEFE) scheme to support efficient model encryption and non-interactive decryption. DEFE also eliminates third-party reliance and defends against server-side inference attacks. Second, we develop a privacy-preserving defensive model aggregation mechanism based on DEFE. This mechanism filters poisonous models under Non-IID data by layer-wise projection and clustering-based analysis. Theoretical analysis and extensive experiments show that SRFed outperforms state-of-the-art baselines in privacy protection, Byzantine robustness, and efficiency.

Recent studies have shown that large language models (LLMs) can infer private user attributes (e.g., age, location, gender) from user-generated text shared online, enabling rapid and large-scale privacy breaches. Existing anonymization-based defenses are coarse-grained, lacking word-level precision in anonymizing privacy-leaking elements. Moreover, they are inherently limited as altering user text to hide sensitive cues still allows attribute inference to occur through models' reasoning capabilities. To address these limitations, we propose a unified defense framework that combines fine-grained anonymization (TRACE) with inference-preventing optimization (RPS). TRACE leverages attention mechanisms and inference chain generation to identify and anonymize privacy-leaking textual elements, while RPS employs a lightweight two-stage optimization strategy to induce model rejection behaviors, thereby preventing attribute inference. Evaluations across diverse LLMs show that TRACE-RPS reduces attribute inference accuracy from around 50\% to below 5\% on open-source models. In addition, our approach offers strong cross-model generalization, prompt-variation robustness, and utility-privacy tradeoffs. Our code is available at https://github.com/Jasper-Yan/TRACE-RPS.

Retrieval-augmented generation (RAG) is increasingly deployed in real-world applications, where its reference-grounded design makes outputs appear trustworthy. This trust has spurred research on poisoning attacks that craft malicious content, inject it into knowledge sources, and manipulate RAG responses. However, when evaluated in practical RAG systems, existing attacks suffer from severely degraded effectiveness. This gap stems from two overlooked realities: (i) content is often processed before use, which can fragment the poison and weaken its effect, and (ii) users often do not issue the exact queries anticipated during attack design. These factors can lead practitioners to underestimate risks and develop a false sense of security. To better characterize the threat to practical systems, we present Confundo, a learning-to-poison framework that fine-tunes a large language model as a poison generator to achieve high effectiveness, robustness, and stealthiness. Confundo provides a unified framework supporting multiple attack objectives, demonstrated by manipulating factual correctness, inducing biased opinions, and triggering hallucinations. By addressing these overlooked challenges, Confundo consistently outperforms a wide range of purpose-built attacks across datasets and RAG configurations by large margins, even in the presence of defenses. Beyond exposing vulnerabilities, we also present a defensive use case that protects web content from unauthorized incorporation into RAG systems via scraping, with no impact on user experience.

Mixture-of-Experts (MoE) has demonstrated strong performance in video understanding tasks, yet its adversarial robustness remains underexplored. Existing attack methods often treat MoE as a unified architecture, overlooking the independent and collaborative weaknesses of key components such as routers and expert modules. To fill this gap, we propose Temporal Lipschitz-Guided Attacks (TLGA) to thoroughly investigate component-level vulnerabilities in video MoE models. We first design attacks on the router, revealing its independent weaknesses. Building on this, we introduce Joint Temporal Lipschitz-Guided Attacks (J-TLGA), which collaboratively perturb both routers and experts. This joint attack significantly amplifies adversarial effects and exposes the Achilles' Heel (collaborative weaknesses) of the MoE architecture. Based on these insights, we further propose Joint Temporal Lipschitz Adversarial Training (J-TLAT). J-TLAT performs joint training to further defend against collaborative weaknesses, enhancing component-wise robustness. Our framework is plug-and-play and reduces inference cost by more than 60% compared with dense models. It consistently enhances adversarial robustness across diverse datasets and architectures, effectively mitigating both the independent and collaborative weaknesses of MoE.

Multilingual safety remains significantly imbalanced, leaving non-high-resource (NHR) languages vulnerable compared to robust high-resource (HR) ones. Moreover, the neural mechanisms driving safety alignment remain unclear despite observed cross-lingual representation transfer. In this paper, we find that LLMs contain a set of cross-lingual shared safety neurons (SS-Neurons), a remarkably small yet critical neuronal subset that jointly regulates safety behavior across languages. We first identify monolingual safety neurons (MS-Neurons) and validate their causal role in safety refusal behavior through targeted activation and suppression. Our cross-lingual analyses then identify SS-Neurons as the subset of MS-Neurons shared between HR and NHR languages, serving as a bridge to transfer safety capabilities from HR to NHR domains. We observe that suppressing these neurons causes concurrent safety drops across NHR languages, whereas reinforcing them improves cross-lingual defensive consistency. Building on these insights, we propose a simple neuron-oriented training strategy that targets SS-Neurons based on language resource distribution and model architecture. Experiments demonstrate that fine-tuning this tiny neuronal subset outperforms state-of-the-art methods, significantly enhancing NHR safety while maintaining the model's general capabilities. The code and dataset will be available athttps://github.com/1518630367/SS-Neuron-Expansion.

Despite their capabilities, large foundation models (LFMs) remain susceptible to adversarial manipulation. Current defenses predominantly rely on the "locality hypothesis", suppressing isolated neurons or features. However, harmful semantics act as distributed, cross-layer circuits, rendering such localized interventions brittle and detrimental to utility. To bridge this gap, we propose \textbf{TraceRouter}, a path-level framework that traces and disconnects the causal propagation circuits of illicit semantics. TraceRouter operates in three stages: (1) it pinpoints a sensitive onset layer by analyzing attention divergence; (2) it leverages sparse autoencoders (SAEs) and differential activation analysis to disentangle and isolate malicious features; and (3) it maps these features to downstream causal pathways via feature influence scores (FIS) derived from zero-out interventions. By selectively suppressing these causal chains, TraceRouter physically severs the flow of harmful information while leaving orthogonal computation routes intact. Extensive experiments demonstrate that TraceRouter significantly outperforms state-of-the-art baselines, achieving a superior trade-off between adversarial robustness and general utility. Our code will be publicly released. WARNING: This paper contains unsafe model responses.

Transformer-based deep neural networks (DNNs) affected by backdoor attacks and unfairness typically exhibit anomalous attention patterns, leading to over-attend to backdoor triggers or protected attributes. Existing neuron-editing mitigation strategies often struggle to handle such situation and most of them lack flexibility and tend to distort feature representations. Motivated by such over-attention phenomenon and software engineering paradigms such as delta debugging and hot patching, we propose AtPatch, a hot-fix method that dynamically redistributes attention maps during model inference. Specifically, for a given input, AtPatch first extracts the attention map from the model's inference process. Then, it uses a pre-trained detector to identify anomalous columns and replace them with unified benign attention. Then, AtPatch rescales other columns to mitigate the impact of over-attention. Finally, AtPatch returns the redistributed attention map to the model for continued inference. Notably, if the detector does not report any anomalous columns, AtPatch directly returns the original attention map to the model. Unlike existing techniques, AtPatch selectively redistributes the attention map, making it better at preserving the model's original functionality. Furthermore, AtPatch's on-the-fly nature allows it to work without modifying model parameters or retraining, making it better suited for deployed models. We conducted extensive experiments to validate AtPatch. Experimental results show that, compared to existing methods, AtPatch can more effectively mitigate backdoor attacks and unfairness while better preserving the model's original functionality.

Despite the intrinsic risk-awareness of Large Language Models (LLMs), current defenses often result in shallow safety alignment, rendering models vulnerable to disguised attacks (e.g., prefilling) while degrading utility. To bridge this gap, we propose SafeThinker, an adaptive framework that dynamically allocates defensive resources via a lightweight gateway classifier. Based on the gateway's risk assessment, inputs are routed through three distinct mechanisms: (i) a Standardized Refusal Mechanism for explicit threats to maximize efficiency; (ii) a Safety-Aware Twin Expert (SATE) module to intercept deceptive attacks masquerading as benign queries; and (iii) a Distribution-Guided Think (DDGT) component that adaptively intervenes during uncertain generation. Experiments show that SafeThinker significantly lowers attack success rates across diverse jailbreak strategies without compromising utility, demonstrating that coordinating intrinsic judgment throughout the generation process effectively balances robustness and practicality.

Large multimodal model powered GUI agents are emerging as high-privilege operators on mobile platforms, entrusted with perceiving screen content and injecting inputs. However, their design operates under the implicit assumption of Visual Atomicity: that the UI state remains invariant between observation and action. We demonstrate that this assumption is fundamentally invalid in Android, creating a critical attack surface. We present Action Rebinding, a novel attack that allows a seemingly-benign app with zero dangerous permissions to rebind an agent's execution. By exploiting the inevitable observation-to-action gap inherent in the agent's reasoning pipeline, the attacker triggers foreground transitions to rebind the agent's planned action toward the target app. We weaponize the agent's task-recovery logic and Android's UI state preservation to orchestrate programmable, multi-step attack chains. Furthermore, we introduce an Intent Alignment Strategy (IAS) that manipulates the agent's reasoning process to rationalize UI states, enabling it to bypass verification gates (e.g., confirmation dialogs) that would otherwise be rejected. We evaluate Action Rebinding Attacks on six widely-used Android GUI agents across 15 tasks. Our results demonstrate a 100% success rate for atomic action rebinding and the ability to reliably orchestrate multi-step attack chains. With IAS, the success rate in bypassing verification gates increases (from 0% to up to 100%). Notably, the attacker application requires no sensitive permissions and contains no privileged API calls, achieving a 0% detection rate across malware scanners (e.g., VirusTotal). Our findings reveal a fundamental architectural flaw in current agent-OS integration and provide critical insights for the secure design of future agent systems. To access experimental logs and demonstration videos, please contact yi_qian@smail.nju.edu.cn.

LLM-based code agents(e.g., ChatGPT Codex) are increasingly deployed as detector for code review and security auditing tasks. Although CoT-enhanced LLM vulnerability detectors are believed to provide improved robustness against obfuscated malicious code, we find that their reasoning chains and semantic abstraction processes exhibit exploitable systematic weaknesses.This allows attackers to covertly embed malicious logic, bypass code review, and propagate backdoored components throughout real-world software supply chains.To investigate this issue, we present CoTDeceptor, the first adversarial code obfuscation framework targeting CoT-enhanced LLM detectors. CoTDeceptor autonomously constructs evolving, hard-to-reverse multi-stage obfuscation strategy chains that effectively disrupt CoT-driven detection logic.We obtained malicious code provided by security enterprise, experimental results demonstrate that CoTDeceptor achieves stable and transferable evasion performance against state-of-the-art LLMs and vulnerability detection agents. CoTDeceptor bypasses 14 out of 15 vulnerability categories, compared to only 2 bypassed by prior methods. Our findings highlight potential risks in real-world software supply chains and underscore the need for more robust and interpretable LLM-powered security analysis systems.

Deep neural networks are widely deployed with quantization techniques to reduce memory and computational costs by lowering the numerical precision of their parameters. While quantization alters model parameters and their outputs, existing privacy analyses primarily focus on full-precision models, leaving a gap in understanding how bit-width reduction can affect privacy leakage. We present the first systematic study of the privacy-utility relationship in post-training quantization (PTQ), a versatile family of methods that can be applied to pretrained models without further training. Using membership inference attacks as our evaluation framework, we analyze three popular PTQ algorithms-AdaRound, BRECQ, and OBC-across multiple precision levels (4-bit, 2-bit, and 1.58-bit) on CIFAR-10, CIFAR-100, and TinyImageNet datasets. Our findings consistently show that low-precision PTQs can reduce privacy leakage. In particular, lower-precision models demonstrate up to an order of magnitude reduction in membership inference vulnerability compared to their full-precision counterparts, albeit at the cost of decreased utility. Additional ablation studies on the 1.58-bit quantization level show that quantizing only the last layer at higher precision enables fine-grained control over the privacy-utility trade-off. These results offer actionable insights for practitioners to balance efficiency, utility, and privacy protection in real-world deployments.

Recommender systems have been widely deployed across various domains such as e-commerce and social media, and intelligently suggest items like products and potential friends to users based on their preferences and interaction history, which are often privacy-sensitive. Recent studies have revealed that recommender systems are prone to membership inference attacks (MIAs), where an attacker aims to infer whether or not a user's data has been used for training a target recommender system. However, existing MIAs fail to exploit the unique characteristic of recommender systems, and therefore are only applicable to mixed recommender systems consisting of two recommendation algorithms. This leaves a gap in investigating MIAs against hybrid-based recommender systems where the same algorithm utilizing user-item historical interactions and attributes of users and items serves and produces personalised recommendations. To investigate how the personalisation in hybrid-based recommender systems influences MIA, we propose a novel metric-based MIA. Specifically, we leverage the characteristic of personalisation to obtain reference recommendation for any target users. Then, a relative membership metric is proposed to exploit a target user's historical interactions, target recommendation, and reference recommendation to infer the membership of the target user's data. Finally, we theoretically and empirically demonstrate the efficacy of the proposed metric-based MIA on hybrid-based recommender systems.