Reinforcement learning with verifiable rewards (RLVR) is central to training modern reasoning models, but the undisclosed training data raises concerns about benchmark contamination. Unlike pretraining methods, which optimize models using token-level probabilities, RLVR fine-tunes models based on reward feedback from self-generated reasoning trajectories, making conventional likelihood-based detection methods less effective. We show that RLVR induces a distinctive behavioral signature: prompts encountered during RLVR training result in more rigid and similar generations, while unseen prompts retain greater diversity. We introduce Min-$k$NN Distance, a simple black-box detector that quantifies this collapse by sampling multiple completions for a given prompt and computing the average of the $k$ smallest nearest-neighbor edit distances. Min-$k$NN Distance requires no access to the reference model or token probabilities. Experiments across multiple RLVR-trained reasoning models show that Min-$k$NN Distance reliably distinguishes RL-seen examples from unseen ones and outperforms existing membership inference and RL contamination detection baselines.

Zero-shot methods detect LLM-generated text by computing statistical signatures using a surrogate model. Existing approaches typically employ a fixed surrogate for all inputs regardless of the unknown source. We systematically examine this design and find that detection performance varies substantially depending on surrogate-source alignment. We observe that while no single surrogate achieves optimal performance universally, a well-matched surrogate typically exists within a diverse pool for any given input. This finding transforms robust detection into a routing problem: selecting the most appropriate surrogate for each input. We propose DetectRouter, a prototype-based framework that learns text-detector affinity through two-stage training. The first stage constructs discriminative prototypes from white-box models; the second generalizes to black-box sources by aligning geometric distances with observed detection scores. Experiments on EvoBench and MAGE benchmarks demonstrate consistent improvements across multiple detection criteria and model families.

Zero-shot detection methods for AI-generated text typically aggregate token-level statistics across entire sequences, overlooking the temporal dynamics inherent to autoregressive generation. We analyze over 120k text samples and reveal Late-Stage Volatility Decay: AI-generated text exhibits rapidly stabilizing log probability fluctuations as generation progresses, while human writing maintains higher variability throughout. This divergence peaks in the second half of sequences, where AI-generated text shows 24--32\% lower volatility. Based on this finding, we propose two simple features: Derivative Dispersion and Local Volatility, which computed exclusively from late-stage statistics. Without perturbation sampling or additional model access, our method achieves state-of-the-art performance on EvoBench and MAGE benchmarks and demonstrates strong complementarity with existing global methods.

The widespread deployment of large language models (LLMs) has raised growing concerns about their misuse risks and associated safety issues. While prior studies have examined the safety of LLMs in general usage, code generation, and agent-based applications, their vulnerabilities in automated algorithm design remain underexplored. To fill this gap, this study investigates this overlooked safety vulnerability, with a particular focus on intelligent optimization algorithm design, given its prevalent use in complex decision-making scenarios. We introduce MalOptBench, a benchmark consisting of 60 malicious optimization algorithm requests, and propose MOBjailbreak, a jailbreak method tailored for this scenario. Through extensive evaluation of 13 mainstream LLMs including the latest GPT-5 and DeepSeek-V3.1, we reveal that most models remain highly susceptible to such attacks, with an average attack success rate of 83.59% and an average harmfulness score of 4.28 out of 5 on original harmful prompts, and near-complete failure under MOBjailbreak. Furthermore, we assess state-of-the-art plug-and-play defenses that can be applied to closed-source models, and find that they are only marginally effective against MOBjailbreak and prone to exaggerated safety behaviors. These findings highlight the urgent need for stronger alignment techniques to safeguard LLMs against misuse in algorithm design.

Zero-shot Learning (ZSL) aims to enable image classifiers to recognize images from unseen classes that were not included during training. Unlike traditional supervised classification, ZSL typically relies on learning a mapping from visual features to predefined, human-understandable class concepts. While ZSL models promise to improve generalization and interpretability, their robustness under systematic input perturbations remain unclear. In this study, we present an empirical analysis about the robustness of existing ZSL methods at both classlevel and concept-level. Specifically, we successfully disrupted their class prediction by the well-known non-target class attack (clsA). However, in the Generalized Zero-shot Learning (GZSL) setting, we observe that the success of clsA is only at the original best-calibrated point. After the attack, the optimal bestcalibration point shifts, and ZSL models maintain relatively strong performance at other calibration points, indicating that clsA results in a spurious attack success in the GZSL. To address this, we propose the Class-Bias Enhanced Attack (CBEA), which completely eliminates GZSL accuracy across all calibrated points by enhancing the gap between seen and unseen class probabilities.Next, at concept-level attack, we introduce two novel attack modes: Class-Preserving Concept Attack (CPconA) and NonClass-Preserving Concept Attack (NCPconA). Our extensive experiments evaluate three typical ZSL models across various architectures from the past three years and reveal that ZSL models are vulnerable not only to the traditional class attack but also to concept-based attacks. These attacks allow malicious actors to easily manipulate class predictions by erasing or introducing concepts. Our findings highlight a significant performance gap between existing approaches, emphasizing the need for improved adversarial robustness in current ZSL models.

With the rapid advancement of retrieval-augmented vision-language models, multimodal medical retrieval-augmented generation (MMed-RAG) systems are increasingly adopted in clinical decision support. These systems enhance medical applications by performing cross-modal retrieval to integrate relevant visual and textual evidence for tasks, e.g., report generation and disease diagnosis. However, their complex architecture also introduces underexplored adversarial vulnerabilities, particularly via visual input perturbations. In this paper, we propose Medusa, a novel framework for crafting cross-modal transferable adversarial attacks on MMed-RAG systems under a black-box setting. Specifically, Medusa formulates the attack as a perturbation optimization problem, leveraging a multi-positive InfoNCE loss (MPIL) to align adversarial visual embeddings with medically plausible but malicious textual targets, thereby hijacking the retrieval process. To enhance transferability, we adopt a surrogate model ensemble and design a dual-loop optimization strategy augmented with invariant risk minimization (IRM). Extensive experiments on two real-world medical tasks, including medical report generation and disease diagnosis, demonstrate that Medusa achieves over 90% average attack success rate across various generation models and retrievers under appropriate parameter configuration, while remaining robust against four mainstream defenses, outperforming state-of-the-art baselines. Our results reveal critical vulnerabilities in the MMed-RAG systems and highlight the necessity of robustness benchmarking in safety-critical medical applications. The code and data are available at https://anonymous.4open.science/r/MMed-RAG-Attack-F05A.

Vision-Language-Action models (VLAs) have recently demonstrated remarkable progress in embodied environments, enabling robots to perceive, reason, and act through unified multimodal understanding. Despite their impressive capabilities, the adversarial robustness of these systems remains largely unexplored, especially under realistic multimodal and black-box conditions. Existing studies mainly focus on single-modality perturbations and overlook the cross-modal misalignment that fundamentally affects embodied reasoning and decision-making. In this paper, we introduce VLA-Fool, a comprehensive study of multimodal adversarial robustness in embodied VLA models under both white-box and black-box settings. VLA-Fool unifies three levels of multimodal adversarial attacks: (1) textual perturbations through gradient-based and prompt-based manipulations, (2) visual perturbations via patch and noise distortions, and (3) cross-modal misalignment attacks that intentionally disrupt the semantic correspondence between perception and instruction. We further incorporate a VLA-aware semantic space into linguistic prompts, developing the first automatically crafted and semantically guided prompting framework. Experiments on the LIBERO benchmark using a fine-tuned OpenVLA model reveal that even minor multimodal perturbations can cause significant behavioral deviations, demonstrating the fragility of embodied multimodal alignment.

Federated learning allows multiple participants to collaboratively train a central model without sharing their private data. However, this distributed nature also exposes new attack surfaces. In particular, backdoor attacks allow attackers to implant malicious behaviors into the global model while maintaining high accuracy on benign inputs. Existing attacks usually rely on fixed patterns or adversarial perturbations as triggers, which tightly couple the main and backdoor tasks. This coupling makes them vulnerable to dilution by honest updates and limits their persistence under federated defenses. In this work, we propose an approach to decouple the backdoor task from the main task by dynamically optimizing the backdoor trigger within a min-max framework. The inner layer maximizes the performance gap between poisoned and benign samples, ensuring that the contributions of benign users have minimal impact on the backdoor. The outer process injects the adaptive triggers into the local model. We evaluate our method on both computer vision and natural language tasks, and compare it with six backdoor attack methods under six defense algorithms. Experimental results show that our method achieves good attack performance and can be easily integrated into existing backdoor attack techniques.

Large language model (LLM) agents are widely deployed in real-world applications, where they leverage tools to retrieve and manipulate external data for complex tasks. However, when interacting with untrusted data sources (e.g., fetching information from public websites), tool responses may contain injected instructions that covertly influence agent behaviors and lead to malicious outcomes, a threat referred to as Indirect Prompt Injection (IPI). Existing defenses typically rely on advanced prompting strategies or auxiliary detection models. While these methods have demonstrated some effectiveness, they fundamentally rely on assumptions about the model's inherent security, which lacks structural constraints on agent behaviors. As a result, agents still retain unrestricted access to tool invocations, leaving them vulnerable to stronger attack vectors that can bypass the security guardrails of the model. To prevent malicious tool invocations at the source, we propose a novel defensive task execution paradigm, called IPIGuard, which models the agents' task execution process as a traversal over a planned Tool Dependency Graph (TDG). By explicitly decoupling action planning from interaction with external data, IPIGuard significantly reduces unintended tool invocations triggered by injected instructions, thereby enhancing robustness against IPI attacks. Experiments on the AgentDojo benchmark show that IPIGuard achieves a superior balance between effectiveness and robustness, paving the way for the development of safer agentic systems in dynamic environments.

The field of AI-generated text detection has evolved from supervised classification to zero-shot statistical analysis. However, current approaches share a fundamental limitation: they aggregate token-level measurements into scalar scores, discarding positional information about where anomalies occur. Our empirical analysis reveals that AI-generated text exhibits significant non-stationarity, statistical properties vary by 73.8\% more between text segments compared to human writing. This discovery explains why existing detectors fail against localized adversarial perturbations that exploit this overlooked characteristic. We introduce Temporal Discrepancy Tomography (TDT), a novel detection paradigm that preserves positional information by reformulating detection as a signal processing task. TDT treats token-level discrepancies as a time-series signal and applies Continuous Wavelet Transform to generate a two-dimensional time-scale representation, capturing both the location and linguistic scale of statistical anomalies. On the RAID benchmark, TDT achieves 0.855 AUROC (7.1\% improvement over the best baseline). More importantly, TDT demonstrates robust performance on adversarial tasks, with 14.1\% AUROC improvement on HART Level 2 paraphrasing attacks. Despite its sophisticated analysis, TDT maintains practical efficiency with only 13\% computational overhead. Our work establishes non-stationarity as a fundamental characteristic of AI-generated text and demonstrates that preserving temporal dynamics is essential for robust detection.