Large Language Model (LLM) watermarking is crucial for establishing the provenance of machine-generated text, but most existing methods rely on a centralized trust model. This model forces users to reveal potentially sensitive text to a provider for detection and offers no way to verify the integrity of the result. While asymmetric schemes have been proposed to address these issues, they are either impractical for short texts or lack formal guarantees linking watermark insertion and detection. We propose VOW, a new protocol that achieves both privacy-preserving and cryptographically verifiable watermark detection with high efficiency. Our approach formulates detection as a secure two-party computation problem, instantiating the watermark's core logic with a Verifiable Oblivious Pseudorandom Function (VOPRF). This allows the user and provider to perform detection without the user's text being revealed, while the provider's result is verifiable. Our comprehensive evaluation shows that VOW is practical for short texts and provides a crucial reassessment of watermark robustness against modern paraphrasing attacks.

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.

Perturbation probing generates task-specific causal hypotheses for FFN neurons in large language models using two forward passes per prompt and no backpropagation, followed by a one-time intervention sweep of about 150 passes amortized across all identified neurons. Across eight behavioral circuits, 13 models, and four architecture families, we identify two circuit structures that organize LLM behavior. Opposition circuits appear when RLHF suppresses a pre-training tendency. In safety refusal, about 50 neurons, or 0.014 percent of all neurons, control the refusal template; ablating them changes 80 percent of response formats on 520 AdvBench prompts while producing near-zero harmful compliance, 3 of 520 cases, all with disclaimers. Routing circuits appear for pre-training behaviors distributed through attention. For language selection, residual-stream direction injection switches English to Chinese output on 99.1 percent of 580 benchmark prompts in the 3 of 19 tested models that satisfy three observed conditions: bilingual training, FFN-to-skip signal ratio between 0.3 and 1.1, and linear representability. The same intervention fails on the other 16 models and on math, code, and factual circuits, defining the limits of directional steering. The FFN-to-skip signal ratio, computed from the same two forward passes, distinguishes the two structures and predicts the appropriate intervention. Circuit topology varies by architecture, from Qwen's concentrated FFN bottleneck to Gemma's normalization-shielded circuit. In Qwen3.5-2B, ablating 20 neurons eliminates multi-turn sycophantic capitulation, while amplifying 10 related neurons improves factual correction from 52 percent to 88 percent on 200 TruthfulQA prompts. These results show that perturbation probing offers mechanistic insight into RLHF-organized behavior and a practical toolkit for precision template-layer editing.

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.

Decompositional jailbreaks pose a critical threat to large language models (LLMs) by allowing adversaries to fragment a malicious objective into a sequence of individually benign queries that collectively reconstruct prohibited content. In real-world deployments, LLMs face a continuous, untraceable stream of fully anonymized and arbitrarily interleaved requests, infiltrated by covertly distributed adversarial queries. Under this rigorous threat model, state-of-the-art defensive strategies exhibit fundamental limitations. In the absence of trustworthy user metadata, they are incapable of tracking global historical contexts, while their deployment of generative models for real-time monitoring introduces computationally prohibitive overhead. To address this, we present TwinGate, a stateful dual-encoder defense framework. TwinGate employs Asymmetric Contrastive Learning (ACL) to cluster semantically disparate but intent-matched malicious fragments in a shared latent space, while a parallel frozen encoder suppresses false positives arising from benign topical overlap. Each request requires only a single lightweight forward pass, enabling the defense to execute in parallel with the target model's prefill phase at negligible latency overhead. To evaluate our approach and advance future research, we construct a comprehensive dataset of over 3.62 million instructions spanning 8,600 distinct malicious intents. Evaluated on this large-scale corpus under a strictly causal protocol, TwinGate achieves high malicious intent recall at a remarkably low false positive rate while remaining highly robust against adaptive attacks. Furthermore, our proposal substantially outperforms stateful and stateless baselines, delivering superior throughput and reduced latency.

Access to genomic data is highly regulated due to its sensitive nature. While safeguards are essential, cumbersome data access processes pose a significant barrier to the development of AI methods for genomics. Synthetic data generation can mitigate this tension by enabling broader data sharing without exposing sensitive information. Synthetic genomic data are produced by training generative models on real data and subsequently sampling artificial data that preserves relevant statistics while limiting disclosures about the underlying individuals. In some settings, a single data holder may have sufficient data to train such generative models; however, in many applications data must be combined across multiple sites to achieve adequate scale. This need arises, e.g., in rare disease studies, where individual hospitals typically hold data for only a small number of patients. The solution we present in this paper enables multiple data holders to jointly train a synthetic data generator without revealing their raw data. Our approach combines secure multiparty computation (MPC) to ensure input privacy, so that no party ever discloses its data in unencrypted form, with differential privacy (DP) to provide output privacy by mitigating information leakage from the released synthetic data. We empirically demonstrate the effectiveness of the proposed method by generating high-utility synthetic datasets from multiple real RNA-seq cohorts in federated settings, showing that our approach enables privacy-preserving data synthesis even when data are distributed across institutions.

Long-context large language models (LLMs)-for example, Gemini-3.1-Pro and Qwen-3.5-are widely used to empower many real-world applications, such as retrieval-augmented generation, autonomous agents, and AI assistants. However, security remains a major concern for their widespread deployment, with threats such as prompt injection and knowledge corruption. To quantify the security risks faced by LLMs under these threats, the research community has developed heuristic-based and optimization-based red-teaming methods. Optimization-based methods generally produce stronger attacks than heuristic attacks and thus provide a more rigorous assessment of LLM security risks. However, they are often resource-intensive, requiring significant computation and GPU memory, especially for long context scenarios. The resource-intensive nature poses a major obstacle for the community (especially academic researchers) to systematically evaluate the security risks of long-context LLMs and assess the effectiveness of defense strategies at scale. In this work, we propose FlashRT, the first framework to improve the efficiency (in terms of both computation and memory) for optimization-based prompt injection and knowledge corruption attacks under long-context LLMs. Through extensive evaluations, we find that FlashRT consistently delivers a 2x-7x speedup (e.g., reducing runtime from one hour to less than ten minutes) and a 2x-4x reduction in GPU memory consumption (e.g., reducing from 264.1 GB to 65.7 GB GPU memory for a 32K token context) compared to state-of-the-art baseline nanoGCG. FlashRT can be broadly applied to black-box optimization methods, such as TAP and AutoDAN. We hope FlashRT can serve as a red-teaming tool to enable systematic evaluation of long-context LLM security. The code is available at: https://github.com/Wang-Yanting/FlashRT

Mixture-of-Experts (MoE) architectures in Large Language Models (LLMs) have significantly reduced inference costs through sparse activation. However, this sparse activation paradigm also introduces new safety challenges. Since only a subset of experts is engaged for each input, model behavior becomes coupled to routing decisions, yielding a difficult-to-control mechanism that can vary across safety-relevant scenarios. At the same time, adapting model behavior through full fine-tuning or retraining is costly, especially when developers need to rapidly configure the same model for different safety objectives. We present MASCing (MoE Activation Steering Configuration), the first framework that enables flexible reconfiguration of MoE behavior across diverse safety scenarios without retraining. MASCing uses an LSTM-based surrogate model to capture cross-layer routing dependencies and map routing logits to downstream behaviors. It then optimizes a steering matrix to identify behavior-relevant expert circuits and, at inference time, applies steering masks to the routing gates to override expert selection. This enables targeted enhancement or suppression of specific behaviors while preserving general language utility. To demonstrate its reconfigurability, we apply MASCing to two different safety-related objectives and observe consistent gains with negligible overhead across seven open-source MoE models. For multi-turn jailbreak defense, it improves the average defense success rate from 52.5% to 83.9%, with gains of up to 89.2%. For adult-content generation, MASCing enables models to comply with such requests that would otherwise be refused, increasing the average generation success rate from 52.6% to 82.0%, with gains of up to 93.0%. These results establish MASCing as a practical, lightweight, and flexible framework for scenario-specific safety reconfiguration in MoE models.

Federated learning (FL) is a popular distributed learning paradigm in machine learning, which enables multiple clients to collaboratively train models under the guidance of a server without exposing private client data. However, FL's decentralized nature makes it vulnerable to poisoning attacks, where malicious clients can submit corrupted models to manipulate the system. To counter such attacks, although various Byzantine-robust methods have been proposed, these methods struggle to provide balanced defense against multiple types of attacks or rely on possessing the dataset in the server. To deal with these drawbacks, thus, we propose an effective multi-layer defensive adaptive aggregation for Bzantine-robust federated learning (AdaBFL) based on a novel three-layer defensive mechanism, which can adaptively adjust the weights of defense algorithms to counter complex attacks. Moreover, we provide convergence properties of our AdaBFL method under the non-convex setting on non-iid data. Comprehensive experiments across multiple datasets validate the superiority of our AdaBFL over the comparable algorithms.

Current DeepFake detection scenarios are mostly binary, yet data manipulation can vary across audio, video, or both, whose variability is not captured in binary settings. Four-class audio-visual formulations address this by discriminating manipulation type, but introduce a unresolved problem: models may rely solely on data source integrity to detect DeepFakes without evaluating their semantic consistency. If the DeepFake origin is not in the data source but in its content, can semantic mismatch be assessed by the state-of-the-art? This paper proposes a new evaluation setup, extending the four-class formulation by explicitly modeling semantic-level inconsistency between authentic modalities with the introduction a new class: Real Audio-Real Video with Semantic Mismatch (RARV-SMM). We assess the robustness of state-of-the-art models in this new realistic DeepFake setting, using the FakeAVCeleb dataset, highlighting the limitations of existing approaches when faced with semantic mismatch data. We further introduce three RARV-SMM variants that expose distinct architectural vulnerabilities as audio-visual divergence increases. We also propose a semantic reinforcement strategy that incorporates the semantic mismatch class and ImageBind embeddings to improve DeepFake detection in both our proposed and state-of-the-art settings, on FakeAVCeleb and LAV-DF, paving the way to more realistic DeepFake detectors. The source code and data are available at https://github.com/.

AI-generated images are becoming increasingly realistic and diverse, posing significant challenges for generalizable detection. While Vision Foundation Models (VFMs) provide rich semantic representations and frequency-based methods capture complementary artifact cues, existing approaches that combine these modalities still suffer from limited generalization, with notable performance degradation on unseen generative models. We attribute this limitation to two key factors: frequency shortcut bias toward easily distinguishable cues associated with specific generators and cross-domain representation conflict between high-level semantics and low-level frequency patterns. To address these issues, we propose a Frequency-aware Gated Injection Network (FGINet) to improve generalization. Specifically, we design a Band-Masked Frequency Encoder (BMFE) that applies cross-band masking in the frequency domain to reduce reliance on generator-specific patterns and encourage more diverse and generalizable representations. We further introduce a Layer-wise Gated Frequency Injection (LGFI) mechanism to progressively inject frequency cues into the VFM backbone with adaptive gating, aligning with its hierarchical abstraction and alleviating representation conflict. Moreover, we propose a Hyperspherical Compactness Learning (HCL) framework with a cosine margin objective to learn compact and well-separated representations. Extensive experiments demonstrate that FGINet achieves state-of-the-art performance and strong generalization across multiple challenging datasets.

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.

Safety-aligned language models must refuse harmful requests without collapsing into broad over-refusal, but the training-time mechanisms behind this tradeoff remain unclear. Prior work characterizes refusal directions and jailbreak robustness, yet does not explain how dynamic adversarial fine-tuning changes refusal carriers across training. We present a measurement-driven mechanism study, not a new defense, on one 7B backbone under supervised fine-tuning (SFT) and R2D2-style dynamic adversarial fine-tuning. Our protocol aligns fixed-source HarmBench, StrongREJECT, and XSTest with a five-anchor refusal-geometry suite and causal interventions. R2D2 drives fixed-source HarmBench ASR to 0.000 at steps 50 and 100, then partially reopens to 0.035 at step 250 and 0.250 at step 500; SFT remains less robust, with ASR between 0.505 and 0.588 at the same anchors. On XSTest, R2D2 any-refusal is 1.000 early, then falls to 0.664 and 0.228. Geometrically, R2D2 preserves a late-layer admissible carrier through step 100 before relocating to an early-layer carrier, while effective rank remains near 1.23--1.27. Causal interventions indicate low-dimensional but utility-coupled control. These results support a reorganization account rather than a drift-only account, with evidence limited to one backbone and fixed-source attacks.

Adversarial Training (AT) is one of the most effective methods for developing robust deep neural networks (DNNs). However, AT faces a trade-off problem between clean accuracy and adversarial robustness. In this work, we reveal a surprising phenomenon for the first time: Varying input perturbation intensities for training samples near decision boundaries in AT have minimal impact on model robustness. This finding directly exposes the inconsistency between accuracy and robustness score fluctuations, leading us to identify the misalignment between input and latent spaces as a critical driver of the robustness-accuracy trade-off. To mitigate this misalignment for harmonizing accuracy and robustness, we define Robust Alignment as a new AT target, encouraging the model perception to change with input perturbations provided the final label prediction remains unchanged, which can be achieved via two novel ideas. First, we suggest a reduced and fixed perturbation intensity for those boundary samples, which facilitates the model to utilize the perturbations as learnable patterns, instead of noises that complicate decision boundaries meaninglessly. Second, we propose a Domain Interpolation Consistency Adversarial Regularization (DICAR), based on rigorous theoretical derivations, which explicitly introduces semantic alignment between input and latent spaces into AT. Based on these two ideas, we end up with a new Robust Alignment Adversarial Training (RAAT) method, effectively harmonizing accuracy and robustness. Extensive experiments on CIFAR-10, CIFAR-100, and Tiny-ImageNet with ResNet-18, PreActResNet-18, and WideResNet-28-10 demonstrate the effectiveness of RAAT in improving the trade-off beyond four common baselines and a total of 14 related state-of-the-art (SOTA) works.

Alignment faking (AF) occurs when an LLM strategically complies with training objectives to avoid value modification, reverting to prior preferences once monitoring is lifted. Current detection methods focus on conversational settings and rely primarily on Chain-of-Thought (CoT) analysis, which provides a reliable signal when strategic reasoning surfaces, but cannot distinguish deception from capability failures if traces are absent or unfaithful. We formalize AF as a composite behavioural event and detect it through observable tool selection, where the LLM selects the safe tool when unmonitored, but switches to the unsafe tool under monitoring that rewards helpfulness over safety, while its reasoning still acknowledges the safe choice. We release a dataset of 108 enterprise IT scenarios spanning Security, Privacy, and Integrity domains under Corruption and Sabotage pressures. Evaluating six frontier LLMs across five independent runs, we find mean AF detection rates between 3.5% and 23.7%, with vulnerability profiles varying by domain and pressure type. These results suggest that susceptibility reflects training methodology rather than capability alone.

Unlike single-face forgeries, deepfakes in complex multi-person interaction scenarios (such as group photos and multi-person meetings) more closely reflect real-world threats. Although existing proactive forensics solutions demonstrate good performance, they heavily rely on a "single-face" setting, making it difficult to effectively address the problems of deepfake localization and source tracing in complex multi-person environments. To address this challenge, we propose the Deep Attributable Watermarking Framework (DAWF). This framework adopts a novel multi-face encoder-decoder architecture that bypasses the cumbersome offline pre-processing steps of traditional forensics, facilitating efficient in-network parallel watermark embedding and cross-face collaborative processing. Crucially, we propose a selective regional supervision loss. This innovative mechanism guides the decoder to focus exclusively on the facial regions tampered with by deepfakes. Leveraging this mechanism alongside the embedded identity payloads, DAWF realizes the "which + who" goal, answering the dual questions of which facial region was forged and who was forged. Extensive experiments on challenging multi-face datasets show that DAWF achieves excellent deepfake localization and traceability in complex multi-person scenes.

Dynamic quantization emerged as a practical approach to increase the utilization and efficiency of the machine learning serving flow. Unlike static quantization, which applies quantization offline, dynamic quantization operates on tensors at run-time, adapting its parameters to the actual input data. Today's mainstream machine learning frameworks, including ML compilers and inference engines, frequently recommend dynamic quantization as an initial step for optimizing model serving. This is because dynamic quantization can significantly reduce memory usage and computational load, leading to faster token generation and improved model serving efficiency without substantial loss in model accuracy. In this paper, we reveal a critical vulnerability in dynamic quantization: an adversary can exploit such quantization strategy to steal sensitive user data placed in the same batch as the adversary's input. Our analysis demonstrates that dynamic quantization, when improperly implemented or configured, can create side channels that expose information about other inputs within the same batch. We call this phenomenon Quantamination, describing contamination from quantization. Specifically, we show that at least 4 of the most popular ML frameworks in use today either default to or can use configurations that leak data across the batch boundary. This data leakage, in theory, allows attackers to partially or even fully recover other users' batched input data, representing a serious privacy risk for existing ML serving frameworks.

Recent methods demonstrate that large-scale pretrained models, such as CLIP vision transformers, effectively detect AI-generated images (AIGIs) from unseen generative models when used as feature extractors. Many state-of-the-art methods for AI-generated image detection build upon the original CLIP-ViT to enhance this generalization. Since CLIP's release, numerous vision foundation models (VFMs) have emerged, incorporating architectural improvements and different training paradigms. Despite these advances, their potential for AIGI detection and AI image forensics remains largely unexplored. In this work, we present a comprehensive benchmark across multiple VFM families, covering diverse pretraining objectives, input resolutions, and model scales. We systematically evaluate their out-of-the-box performance for detecting fully-generated AI-images and AI-inpainted images, and discover that the best model outperforms the original CLIP by more than 12% in accuracy, beating established approaches in the process. To fully leverage the features of a modern VFM, we propose a simple redesign of the classifier head by utilizing tunable attention pooling (TAP), which aggregates output tokens into a refined global representation. Integrating TAP with the latest VFMs yields substantial performance gains across several AIGI detection benchmarks, establishing a new state-of-the-art on two challenging benchmarks for in-the-wild detection of AI-generated and -inpainted images.

As Large Language Models (LLMs) are increasingly integrated into academic peer review, their vulnerability to adversarial prompts -- adversarial instructions embedded in submissions to manipulate outcomes -- emerges as a critical threat to scholarly integrity. To counter this, we propose a novel adversarial framework where a Generator model, trained to create sophisticated attack prompts, is jointly optimized with a Defender model tasked with their detection. This system is trained using a loss function inspired by Information Retrieval Generative Adversarial Networks, which fosters a dynamic co-evolution between the two models, forcing the Defender to develop robust capabilities against continuously improving attack strategies. The resulting framework demonstrates significantly enhanced resilience to novel and evolving threats compared to static defenses, thereby establishing a critical foundation for securing the integrity of peer review.

The rapid advancement of large language models (LLMs) presents new security challenges, particularly in detecting machine-generated text used for misinformation, impersonation, and content forgery. Most existing detection approaches struggle with robustness against adversarial perturbation, paraphrasing attacks, and domain shifts, often requiring restrictive access to model parameters or large labeled datasets. To address this, we propose DSIPA, a novel training-free framework that detects LLM-generated content by quantifying sentiment distributional stability under controlled stylistic variation. It is based on the observation that LLMs typically exhibit more emotionally consistent outputs, while human-written texts display greater affective variation. Our framework operates in a zero-shot, black-box manner, leveraging two unsupervised metrics, sentiment distribution consistency and sentiment distribution preservation, to capture these intrinsic behavioral asymmetries without the need for parameter updates or probability access. Extensive experiments are conducted on state-of-the-art proprietary and open-source models, including GPT-5.2, Gemini-1.5-pro, Claude-3, and LLaMa-3.3. Evaluations on five domains, such as news articles, programming code, student essays, academic papers, and community comments, demonstrate that DSIPA improves F1 detection scores by up to 49.89% over baseline methods. The framework exhibits superior generalizability across domains and strong resilience to adversarial conditions, providing a robust and interpretable behavioral signal for secure content identification in the evolving LLM landscape.