Papers

Multi-agent systems powered by large language models are advancing rapidly, yet the tension between mutual trust and security remains underexplored. We introduce and empirically validate the Trust-Vulnerability Paradox (TVP): increasing inter-agent trust to enhance coordination simultaneously expands risks of over-exposure and over-authorization. To investigate this paradox, we construct a scenario-game dataset spanning 3 macro scenes and 19 sub-scenes, and run extensive closed-loop interactions with trust explicitly parameterized. Using Minimum Necessary Information (MNI) as the safety baseline, we propose two unified metrics: Over-Exposure Rate (OER) to detect boundary violations, and Authorization Drift (AD) to capture sensitivity to trust levels. Results across multiple model backends and orchestration frameworks reveal consistent trends: higher trust improves task success but also heightens exposure risks, with heterogeneous trust-to-risk mappings across systems. We further examine defenses such as Sensitive Information Repartitioning and Guardian-Agent enablement, both of which reduce OER and attenuate AD. Overall, this study formalizes TVP, establishes reproducible baselines with unified metrics, and demonstrates that trust must be modeled and scheduled as a first-class security variable in multi-agent system design.

Very few available individual bandwidth reservation schemes provide efficient and cost-effective bandwidth reservation that is required for safety-critical and time-sensitive vehicular networked applications. These schemes allow vehicles to make reservation requests for the required resources. Accordingly, a Mobile Network Operator (MNO) can allocate and guarantee bandwidth resources based on these requests. However, due to uncertainty in future reservation time and bandwidth costs, the design of an optimized reservation strategy is challenging. In this article, we propose a novel multi-objective bandwidth reservation update approach with an optimal strategy based on Double Deep Q-Network (DDQN). The key design objectives are to minimize the reservation cost with multiple MNOs and to ensure reliable resource provisioning in uncertain situations by solving scenarios such as underbooked and overbooked reservations along the driving path. The enhancements and advantages of our proposed strategy have been demonstrated through extensive experimental results when compared to other methods like greedy update or other deep reinforcement learning approaches. Our strategy demonstrates a 40% reduction in bandwidth costs across all investigated scenarios and simultaneously resolves uncertain situations in a cost-effective manner.

LLMs are often used by downstream users as teacher models for knowledge distillation, compressing their capabilities into memory-efficient models. However, as these teacher models may stem from untrusted parties, distillation can raise unexpected security risks. In this paper, we investigate the security implications of knowledge distillation from backdoored teacher models. First, we show that prior backdoors mostly do not transfer onto student models. Our key insight is that this is because existing LLM backdooring methods choose trigger tokens that rarely occur in usual contexts. We argue that this underestimates the security risks of knowledge distillation and introduce a new backdooring technique, T-MTB, that enables the construction and study of transferable backdoors. T-MTB carefully constructs a composite backdoor trigger, made up of several specific tokens that often occur individually in anticipated distillation datasets. As such, the poisoned teacher remains stealthy, while during distillation the individual presence of these tokens provides enough signal for the backdoor to transfer onto the student. Using T-MTB, we demonstrate and extensively study the security risks of transferable backdoors across two attack scenarios, jailbreaking and content modulation, and across four model families of LLMs.

Security analysts face increasing pressure to triage large and complex vulnerability backlogs. Large Language Models (LLMs) offer a potential aid by automating parts of the interpretation process. We evaluate four models (ChatGPT, Claude, Gemini, and DeepSeek) across twelve prompting techniques to interpret semi-structured and unstructured vulnerability information. As a concrete use case, we test each model's ability to predict decision points in the Stakeholder-Specific Vulnerability Categorization (SSVC) framework: Exploitation, Automatable, Technical Impact, and Mission and Wellbeing. Using 384 real-world vulnerabilities from the VulZoo dataset, we issued more than 165,000 queries to assess performance under prompting styles including one-shot, few-shot, and chain-of-thought. We report F1 scores for each SSVC decision point and Cohen's kappa (weighted and unweighted) for the final SSVC decision outcomes. Gemini consistently ranked highest, leading on three of four decision points and yielding the most correct recommendations. Prompting with exemplars generally improved accuracy, although all models struggled on some decision points. Only DeepSeek achieved fair agreement under weighted metrics, and all models tended to over-predict risk. Overall, current LLMs do not replace expert judgment. However, specific LLM and prompt combinations show moderate effectiveness for targeted SSVC decisions. When applied with care, LLMs can support vulnerability prioritization workflows and help security teams respond more efficiently to emerging threats.

Let $\mathbb{F}_q$ be a finite field, and let $F \in \mathbb{F}_q [X]$ be a polynomial with $d = \text{deg} \left( F \right)$ such that $\gcd \left( d, q \right) = 1$. In this paper we prove that the $c$-Boomerang uniformity, $c \neq 0$, of $F$ is bounded by - $d^2$ if $c^2 \neq 1$, - $d \cdot (d - 1)$ if $c = -1$, - $d \cdot (d - 2)$ if $c = 1$. For all cases of $c$, we present tight examples for $F \in \mathbb{F}_q [X]$. Additionally, for the proof of $c = 1$ we establish that the bivariate polynomial $F (x) - F (y) + a \in k [x, y]$, where $k$ is a field of characteristic $p$ and $a \in k \setminus \{ 0 \}$, is absolutely irreducible if $p \nmid \text{deg} \left( F \right)$.

Phone scams remain a pervasive threat to both personal safety and financial security worldwide. Recent advances in large language models (LLMs) have demonstrated strong potential in detecting fraudulent behavior by analyzing transcribed phone conversations. However, these capabilities introduce notable privacy risks, as such conversations frequently contain sensitive personal information that may be exposed to third-party service providers during processing. In this work, we explore how to harness LLMs for phone scam detection while preserving user privacy. We propose MASK (Modular Adaptive Sanitization Kit), a trainable and extensible framework that enables dynamic privacy adjustment based on individual preferences. MASK provides a pluggable architecture that accommodates diverse sanitization methods - from traditional keyword-based techniques for high-privacy users to sophisticated neural approaches for those prioritizing accuracy. We also discuss potential modeling approaches and loss function designs for future development, enabling the creation of truly personalized, privacy-aware LLM-based detection systems that balance user trust and detection effectiveness, even beyond phone scam context.

This paper explores the challenges of cyberattack attribution, specifically APTs, applying the case study approach for the WhisperGate cyber operation of January 2022 executed by the Russian military intelligence service (GRU) and targeting Ukrainian government entities. The study provides a detailed review of the threat actor identifiers and taxonomies used by leading cybersecurity vendors, focusing on the evolving attribution from Microsoft, ESET, and CrowdStrike researchers. Once the attribution to Ember Bear (GRU Unit 29155) is established through technical and intelligence reports, we use both traditional machine learning classifiers and a large language model (ChatGPT) to analyze the indicators of compromise (IoCs), tactics, and techniques to statistically and semantically attribute the WhisperGate attack. Our findings reveal overlapping indicators with the Sandworm group (GRU Unit 74455) but also strong evidence pointing to Ember Bear, especially when the LLM is fine-tuned or contextually augmented with additional intelligence. Thus, showing how AI/GenAI with proper fine-tuning are capable of solving the attribution challenge.

Large Language Models (LLMs) have shown great promise in automating data analytics tasks by interpreting natural language queries and generating multi-operation execution plans. However, existing LLM-agent-based analytics frameworks operate under the assumption of centralized data access, offering little to no privacy protection. In contrast, federated analytics (FA) enables privacy-preserving computation across distributed data sources, but lacks support for natural language input and requires structured, machine-readable queries. In this work, we present LAFA, the first system that integrates LLM-agent-based data analytics with FA. LAFA introduces a hierarchical multi-agent architecture that accepts natural language queries and transforms them into optimized, executable FA workflows. A coarse-grained planner first decomposes complex queries into sub-queries, while a fine-grained planner maps each subquery into a Directed Acyclic Graph of FA operations using prior structural knowledge. To improve execution efficiency, an optimizer agent rewrites and merges multiple DAGs, eliminating redundant operations and minimizing computational and communicational overhead. Our experiments demonstrate that LAFA consistently outperforms baseline prompting strategies by achieving higher execution plan success rates and reducing resource-intensive FA operations by a substantial margin. This work establishes a practical foundation for privacy-preserving, LLM-driven analytics that supports natural language input in the FA setting.

Web-based behavior-manipulation attacks (BMAs) - such as scareware, fake software downloads, tech support scams, etc. - are a class of social engineering (SE) attacks that exploit human decision-making vulnerabilities. These attacks remain under-studied compared to other attacks such as information harvesting attacks (e.g., phishing) or malware infections. Prior technical work has primarily focused on measuring BMAs, offering little in the way of generic defenses. To address this gap, we introduce Pixel Patrol 3D (PP3D), the first end-to-end browser framework for discovering, detecting, and defending against behavior-manipulating SE attacks in real time. PP3D consists of a visual detection model implemented within a browser extension, which deploys the model client-side to protect users across desktop and mobile devices while preserving privacy. Our evaluation shows that PP3D can achieve above 99% detection rate at 1% false positives, while maintaining good latency and overhead performance across devices. Even when faced with new BMA samples collected months after training the detection model, our defense system can still achieve above 97% detection rate at 1% false positives. These results demonstrate that our framework offers a practical, effective, and generalizable defense against a broad and evolving class of web behavior-manipulation attacks.

The Proof-of-Concept (PoC) for a vulnerability is crucial in validating its existence, mitigating false positives, and illustrating the severity of the security threat it poses. However, research on PoCs significantly lags behind studies focusing on vulnerability data. This discrepancy can be directly attributed to several challenges, including the dispersion of real-world PoCs across multiple platforms, the diversity in writing styles, and the difficulty associated with PoC reproduction. To fill this gap, we conduct the first large-scale study on PoCs in the wild, assessing their report availability, completeness, reproducibility. Specifically, 1) to investigate PoC reports availability for CVE vulnerability, we collected an extensive dataset of 470,921 PoCs and their reports from 13 platforms, representing the broadest collection of publicly available PoCs to date. 2) To assess the completeness of PoC report at a fine-grained level, we proposed a component extraction method, which combines pattern-matching techniques with a fine-tuned BERT-NER model to extract 9 key components from PoC reports. 3) To evaluate the effectiveness of PoCs, we recruited 8 participants to manually reproduce 150 sampled vulnerabilities with 32 vulnerability types based on PoC reports, enabling an in-depth analysis of PoC reproducibility and the factors influencing it. Our findings reveal that 78.9% of CVE vulnerabilities lack available PoCs, and existing PoC reports typically miss about 30% of the essential components required for effective vulnerability understanding and reproduction, with various reasons identified for the failure to reproduce vulnerabilities using available PoC reports. Finally, we proposed actionable strategies for stakeholders to enhance the overall usability of vulnerability PoCs in strengthening software security.

Phishing attacks in Web3 ecosystems are increasingly sophisticated, exploiting deceptive contract logic, malicious frontend scripts, and token approval patterns. We present DeepTx, a real-time transaction analysis system that detects such threats before user confirmation. DeepTx simulates pending transactions, extracts behavior, context, and UI features, and uses multiple large language models (LLMs) to reason about transaction intent. A consensus mechanism with self-reflection ensures robust and explainable decisions. Evaluated on our phishing dataset, DeepTx achieves high precision and recall (demo video: https://youtu.be/4OfK9KCEXUM).

Undoubtedly, the Internet has become one of the most important conduits to information for the general public. Nonetheless, Internet access can be and has been limited systematically or blocked completely during political events in numerous countries and regions by various censorship mechanisms. Depending on where the core filtering component is situated, censorship techniques have been classified as client-based, server-based, or network-based. However, as the Internet evolves rapidly, new and sophisticated censorship techniques have emerged, which involve techniques that cut across locations and involve new forms of hurdles to information access. We argue that modern censorship can be better understood through a new lens that we term chokepoints, which identifies bottlenecks in the content production or delivery cycle where efficient new forms of large-scale client-side surveillance and filtering mechanisms have emerged.

We initiate the study of distribution testing under user-level local differential privacy, where each of $n$ users contributes $m$ samples from the unknown underlying distribution. This setting, albeit very natural, is significantly more challenging that the usual locally private setting, as for the same parameter $\varepsilon$ the privacy guarantee must now apply to a full batch of $m$ data points. While some recent work consider distribution learning in this user-level setting, nothing was known for even the most fundamental testing task, uniformity testing (and its generalization, identity testing). We address this gap, by providing (nearly) sample-optimal user-level LDP algorithms for uniformity and identity testing. Motivated by practical considerations, our main focus is on the private-coin, symmetric setting, which does not require users to share a common random seed nor to have been assigned a globally unique identifier.

Despite progress in watermarking algorithms for large language models (LLMs), real-world deployment remains limited. We argue that this gap stems from misaligned incentives among LLM providers, platforms, and end users, which manifest as three key barriers: competitive risk, detection-tool governance, and attribution issues. We revisit three classes of watermarking through this lens. Model watermarking naturally aligns with LLM provider interests, yet faces new challenges in open-source ecosystems. LLM text watermarking offers modest provider benefit when framed solely as an anti-misuse tool, but can gain traction in narrowly scoped settings such as dataset de-contamination or user-controlled provenance. In-context watermarking (ICW) is tailored for trusted parties, such as conference organizers or educators, who embed hidden watermarking instructions into documents. If a dishonest reviewer or student submits this text to an LLM, the output carries a detectable watermark indicating misuse. This setup aligns incentives: users experience no quality loss, trusted parties gain a detection tool, and LLM providers remain neutral by simply following watermark instructions. We advocate for a broader exploration of incentive-aligned methods, with ICW as an example, in domains where trusted parties need reliable tools to detect misuse. More broadly, we distill design principles for incentive-aligned, domain-specific watermarking and outline future research directions. Our position is that the practical adoption of LLM watermarking requires aligning stakeholder incentives in targeted application domains and fostering active community engagement.

Host-based cryptomining malware, commonly known as cryptojackers, have gained notoriety for their stealth and the significant financial losses they cause in Linux-based cloud environments. Existing solutions often struggle with scalability due to high monitoring overhead, low detection accuracy against obfuscated behavior, and lack of integrated remediation. We present CryptoGuard, a lightweight hybrid solution that combines detection and remediation strategies to counter cryptojackers. To ensure scalability, CryptoGuard uses sketch- and sliding window-based syscall monitoring to collect behavior patterns with minimal overhead. It decomposes the classification task into a two-phase process, leveraging deep learning models to identify suspicious activity with high precision. To counter evasion techniques such as entry point poisoning and PID manipulation, CryptoGuard integrates targeted remediation mechanisms based on eBPF, a modern Linux kernel feature deployable on any compatible host. Evaluated on 123 real-world cryptojacker samples, it achieves average F1-scores of 96.12% and 92.26% across the two phases, and outperforms state-of-the-art baselines in terms of true and false positive rates, while incurring only 0.06% CPU overhead per host.

Generating high-quality synthetic text under differential privacy (DP) is critical for training and evaluating language models without compromising user privacy. Prior work on synthesizing DP datasets often fail to preserve key statistical attributes, suffer utility loss from the noise required by DP, and lack fine-grained control over generation. To address these challenges, we make two contributions. First, we introduce a hierarchical framework that decomposes DP synthetic text generation into two subtasks: feature learning and conditional text generation. This design explicitly incorporates learned features into the generation process and simplifies the end-to-end synthesis task. Through systematic ablations, we identify the most effective configuration: a rich tabular schema as feature, a DP tabular synthesizer, and a DP fine-tuned conditional generator, which we term ACTG (Attribute-Conditioned Text Generation). Second, we propose Anchored RL (ARL), a post-training method that improves the instruction-following ability of ACTG for conditional generation. ARL combines RL to boost control with an SFT anchor on best-of-$N$ data to prevent reward hacking. Together, these components form our end-to-end algorithm ACTG-ARL, which advances both the quality of DP synthetic text (+20% MAUVE over prior work) and the control of the conditional generator under strong privacy guarantees.

Despite recent advances, Large Language Models (LLMs) still generate vulnerable code. Retrieval-Augmented Generation (RAG) has the potential to enhance LLMs for secure code generation by incorporating external security knowledge. However, the conventional RAG design struggles with the noise of raw security-related documents, and existing retrieval methods overlook the significant security semantics implicitly embedded in task descriptions. To address these issues, we propose Rescue, a new RAG framework for secure code generation with two key innovations. First, we propose a hybrid knowledge base construction method that combines LLM-assisted cluster-then-summarize distillation with program slicing, producing both high-level security guidelines and concise, security-focused code examples. Second, we design a hierarchical multi-faceted retrieval that traverses the constructed knowledge base from top to bottom and integrates multiple security-critical facts at each hierarchical level, ensuring comprehensive and accurate retrieval. We evaluated Rescue on four benchmarks and compared it with five state-of-the-art secure code generation methods on six LLMs. The results demonstrate that Rescue improves the SecurePass@1 metric by an average of 4.8 points, establishing a new state-of-the-art performance for security. Furthermore, we performed in-depth analysis and ablation studies to rigorously validate the effectiveness of individual components in Rescue. Our code is available at https://github.com/steven1518/RESCUE.

The inherent determinism of blockchain technology poses a significant challenge to generating secure random numbers within smart contracts, leading to exploitable vulnerabilities, particularly in decentralized finance (DeFi) ecosystems and blockchain-based gaming applications. From our observations, the current state-of-the-art detection tools suffer from inadequate precision while dealing with random number vulnerabilities. To address this problem, we propose TaintSentinel, a novel path sensitive vulnerability detection system designed to analyze smart contracts at the execution path level and gradually analyze taint with domain-specific rules. This paper discusses a solution that incorporates a multi-faceted approach, integrating rule-based taint analysis to track data flow, a dual stream neural network to identify complex vulnerability signatures, and evidence-based parameter initialization to minimize false positives. The system's two-phase operation involves semantic graph construction and taint propagation analysis, followed by pattern recognition using PathGNN and global structural analysis via GlobalGCN. Our experiments on 4,844 contracts demonstrate the superior performance of TaintSentinel relative to existing tools, yielding an F1-score of 0.892, an AUC-ROC of 0.94, and a PRA accuracy of 97%.

As large language models (LLMs) become integrated into various sensitive applications, prompt injection, the use of prompting to induce harmful behaviors from LLMs, poses an ever increasing risk. Prompt injection attacks can cause LLMs to leak sensitive data, spread misinformation, and exhibit harmful behaviors. To defend against these attacks, we propose CourtGuard, a locally-runnable, multiagent prompt injection classifier. In it, prompts are evaluated in a court-like multiagent LLM system, where a "defense attorney" model argues the prompt is benign, a "prosecution attorney" model argues the prompt is a prompt injection, and a "judge" model gives the final classification. CourtGuard has a lower false positive rate than the Direct Detector, an LLM as-a-judge. However, CourtGuard is generally a worse prompt injection detector. Nevertheless, this lower false positive rate highlights the importance of considering both adversarial and benign scenarios for the classification of a prompt. Additionally, the relative performance of CourtGuard in comparison to other prompt injection classifiers advances the use of multiagent systems as a defense against prompt injection attacks. The implementations of CourtGuard and the Direct Detector with full prompts for Gemma-3-12b-it, Llama-3.3-8B, and Phi-4-mini-instruct are available at https://github.com/isaacwu2000/CourtGuard.

The impending adoption of Open Radio Access Network (O-RAN) is fueling innovation in the RAN towards data-driven operation. Unlike traditional RAN where the RAN data and its usage is restricted within proprietary and monolithic RAN equipment, the O-RAN architecture opens up access to RAN data via RAN intelligent controllers (RICs), to third-party machine learning (ML) powered applications - rApps and xApps - to optimize RAN operations. Consequently, a major focus has been placed on leveraging RAN data to unlock greater efficiency gains. However, there is an increasing recognition that RAN data access to apps could become a source of vulnerability and be exploited by malicious actors. Motivated by this, we carry out a comprehensive investigation of data vulnerabilities on both xApps and rApps, respectively hosted in Near- and Non-real-time (RT) RIC components of O-RAN. We qualitatively analyse the O-RAN security mechanisms and limitations for xApps and rApps, and consider a threat model informed by this analysis. We design a viable and effective black-box evasion attack strategy targeting O-RAN RIC Apps while accounting for the stringent timing constraints and attack effectiveness. The strategy employs four key techniques: the model cloning algorithm, input-specific perturbations, universal adversarial perturbations (UAPs), and targeted UAPs. This strategy targets ML models used by both xApps and rApps within the O-RAN system, aiming to degrade network performance. We validate the effectiveness of the designed evasion attack strategy and quantify the scale of performance degradation using a real-world O-RAN testbed and emulation environments. Evaluation is conducted using the Interference Classification xApp and the Power Saving rApp as representatives for near-RT and non-RT RICs. We also show that the attack strategy is effective against prominent defense techniques for adversarial ML.

As users increasingly turn to large language model (LLM) based web agents to automate online tasks, agents may encounter dark patterns: deceptive user interface designs that manipulate users into making unintended decisions. Although dark patterns primarily target human users, their potentially harmful impacts on LLM-based generalist web agents remain unexplored. In this paper, we present the first study that investigates the impact of dark patterns on the decision-making process of LLM-based generalist web agents. To achieve this, we introduce LiteAgent, a lightweight framework that automatically prompts agents to execute tasks while capturing comprehensive logs and screen-recordings of their interactions. We also present TrickyArena, a controlled environment comprising web applications from domains such as e-commerce, streaming services, and news platforms, each containing diverse and realistic dark patterns that can be selectively enabled or disabled. Using LiteAgent and TrickyArena, we conduct multiple experiments to assess the impact of both individual and combined dark patterns on web agent behavior. We evaluate six popular LLM-based generalist web agents across three LLMs and discover that when there is a single dark pattern present, agents are susceptible to it an average of 41% of the time. We also find that modifying dark pattern UI attributes through visual design changes or HTML code adjustments and introducing multiple dark patterns simultaneously can influence agent susceptibility. This study emphasizes the need for holistic defense mechanisms in web agents, encompassing both agent-specific protections and broader web safety measures.

Evaluating the usefulness of data before purchase is essential when obtaining data for high-quality machine learning models, yet both model builders and data providers are often unwilling to reveal their proprietary assets. We present PrivaDE, a privacy-preserving protocol that allows a model owner and a data owner to jointly compute a utility score for a candidate dataset without fully exposing model parameters, raw features, or labels. PrivaDE provides strong security against malicious behavior and can be integrated into blockchain-based marketplaces, where smart contracts enforce fair execution and payment. To make the protocol practical, we propose optimizations to enable efficient secure model inference, and a model-agnostic scoring method that uses only a small, representative subset of the data while still reflecting its impact on downstream training. Evaluation shows that PrivaDE performs data evaluation effectively, achieving online runtimes within 15 minutes even for models with millions of parameters. Our work lays the foundation for fair and automated data marketplaces in decentralized machine learning ecosystems.

The integration of Unmanned Aerial Vehicles (UAVs) into Open Radio Access Networks (O-RAN) enhances communication in disaster management and Search and Rescue (SAR) operations by ensuring connectivity when infrastructure fails. However, SAR scenarios demand stringent security and low-latency communication, as delays or breaches can compromise mission success. While UAVs serve as mobile relays, they introduce challenges in energy consumption and resource management, necessitating intelligent allocation strategies. Existing UAV-assisted O-RAN approaches often overlook the joint optimization of security, latency, and energy efficiency in dynamic environments. This paper proposes a novel Reinforcement Learning (RL)-based framework for dynamic resource allocation in UAV relays, explicitly addressing these trade-offs. Our approach formulates an optimization problem that integrates security-aware resource allocation, latency minimization, and energy efficiency, which is solved using RL. Unlike heuristic or static methods, our framework adapts in real-time to network dynamics, ensuring robust communication. Simulations demonstrate superior performance compared to heuristic baselines, achieving enhanced security and energy efficiency while maintaining ultra-low latency in SAR scenarios.

The convergence of LLM-powered research assistants and AI-based peer review systems creates a critical vulnerability: fully automated publication loops where AI-generated research is evaluated by AI reviewers without human oversight. We investigate this through BadScientist, a framework that evaluates whether fabrication-oriented paper generation agents can deceive multi-model LLM review systems. Our generator employs presentation-manipulation strategies requiring no real experiments. We develop a rigorous evaluation framework with formal error guarantees (concentration bounds and calibration analysis), calibrated on real data. Our results reveal systematic vulnerabilities: fabricated papers achieve acceptance rates up to . Critically, we identify concern-acceptance conflict – reviewers frequently flag integrity issues yet assign acceptance-level scores. Our mitigation strategies show only marginal improvements, with detection accuracy barely exceeding random chance. Despite provably sound aggregation mathematics, integrity checking systematically fails, exposing fundamental limitations in current AI-driven review systems and underscoring the urgent need for defense-in-depth safeguards in scientific publishing.

Large Language Models (LLMs) are improving at an exceptional rate. With the advent of agentic workflows, multi-turn dialogue has become the de facto mode of interaction with LLMs for completing long and complex tasks. While LLM capabilities continue to improve, they remain increasingly susceptible to jailbreaking, especially in multi-turn scenarios where harmful intent can be subtly injected across the conversation to produce nefarious outcomes. While single-turn attacks have been extensively explored, adaptability, efficiency and effectiveness continue to remain key challenges for their multi-turn counterparts. To address these gaps, we present PLAGUE, a novel plug-and-play framework for designing multi-turn attacks inspired by lifelong-learning agents. PLAGUE dissects the lifetime of a multi-turn attack into three carefully designed phases (Primer, Planner and Finisher) that enable a systematic and information-rich exploration of the multi-turn attack family. Evaluations show that red-teaming agents designed using PLAGUE achieve state-of-the-art jailbreaking results, improving attack success rates (ASR) by more than 30% across leading models in a lesser or comparable query budget. Particularly, PLAGUE enables an ASR (based on StrongReject) of 81.4% on OpenAI's o3 and 67.3% on Claude's Opus 4.1, two models that are considered highly resistant to jailbreaks in safety literature. Our work offers tools and insights to understand the importance of plan initialization, context optimization and lifelong learning in crafting multi-turn attacks for a comprehensive model vulnerability evaluation.

Smart contracts play a significant role in automating blockchain services. Nevertheless, vulnerabilities in smart contracts pose serious threats to blockchain security. Currently, traditional detection methods primarily rely on static analysis and formal verification, which can result in high false-positive rates and poor scalability. Large Language Models (LLMs) have recently made significant progress in smart contract vulnerability detection. However, they still face challenges such as high inference costs and substantial computational overhead. In this paper, we propose ParaVul, a parallel LLM and retrieval-augmented framework to improve the reliability and accuracy of smart contract vulnerability detection. Specifically, we first develop Sparse Low-Rank Adaptation (SLoRA) for LLM fine-tuning. SLoRA introduces sparsification by incorporating a sparse matrix into quantized LoRA-based LLMs, thereby reducing computational overhead and resource requirements while enhancing their ability to understand vulnerability-related issues. We then construct a vulnerability contract dataset and develop a hybrid Retrieval-Augmented Generation (RAG) system that integrates dense retrieval with Best Matching 25 (BM25), assisting in verifying the results generated by the LLM. Furthermore, we propose a meta-learning model to fuse the outputs of the RAG system and the LLM, thereby generating the final detection results. After completing vulnerability detection, we design chain-of-thought prompts to guide LLMs to generate comprehensive vulnerability detection reports. Simulation results demonstrate the superiority of ParaVul, especially in terms of F1 scores, achieving 0.9398 for single-label detection and 0.9330 for multi-label detection.

Data unlearning aims to remove the influence of specific training samples from a trained model. In fine-tuning methods, data unlearning relies primarily on loss maximization over forget samples, which often leads to quality degradation or incomplete forgetting. Existing methods perform unlearning uniformly across diffusion stages, ignoring diffusion dynamics from noise to data. Our systematic study of diffusion phases shows that forgetting in diffusion models is uneven across time and frequency, with theoretical justification of distributive distortion and forgetting-utility trade-off. By selectively forgetting time and frequency in diffusion models, we achieve both higher unlearning success rates and improved generation quality across diverse settings, including both conditional and unconditional scenarios. We also introduce an improved SSCD metric that measures dissimilarity using a normalized perturbation distance. Together, we provide practical insights for understanding and improving data unlearning in diffusion models.

Vision-Language Models (VLMs) extend large language models with visual reasoning, but their multimodal design also introduces new, underexplored vulnerabilities. Existing multimodal red-teaming methods largely rely on brittle templates, focus on single-attack settings, and expose only a narrow subset of vulnerabilities. To address these limitations, we introduce VERA-V, a variational inference framework that recasts multimodal jailbreak discovery as learning a joint posterior distribution over paired text-image prompts. This probabilistic view enables the generation of stealthy, coupled adversarial inputs that bypass model guardrails. We train a lightweight attacker to approximate the posterior, allowing efficient sampling of diverse jailbreaks and providing distributional insights into vulnerabilities. VERA-V further integrates three complementary strategies: (i) typography-based text prompts that embed harmful cues, (ii) diffusion-based image synthesis that introduces adversarial signals, and (iii) structured distractors to fragment VLM attention. Experiments on HarmBench and HADES benchmarks show that VERA-V consistently outperforms state-of-the-art baselines on both open-source and frontier VLMs, achieving up to 53.75% higher attack success rate (ASR) over the best baseline on GPT-4o. We include the code on the project page available here: https://github.com/kxwhiowo/VERA-V

Multimodal Large Language Models (MLLMs) achieve strong reasoning and perception capabilities but are increasingly vulnerable to jailbreak attacks. While existing work focuses on explicit attacks, where malicious content resides in a single modality, recent studies reveal implicit attacks, in which benign text and image inputs jointly express unsafe intent. Such joint-modal threats are difficult to detect and remain underexplored, largely due to the scarcity of high-quality implicit data. We propose ImpForge, an automated red-teaming pipeline that leverages reinforcement learning with tailored reward modules to generate diverse implicit samples across 14 domains. Building on this dataset, we further develop CrossGuard, an intent-aware safeguard providing robust and comprehensive defense against both explicit and implicit threats. Extensive experiments across safe and unsafe benchmarks, implicit and explicit attacks, and multiple out-of-domain settings demonstrate that CrossGuard significantly outperforms existing defenses, including advanced MLLMs and guardrails, achieving stronger security while maintaining high utility. This offers a balanced and practical solution for enhancing MLLM robustness against real-world multimodal threats. Our code is released: https://github.com/ZhangXu0963/CrossGuard.

Large Audio-Language Models (LALMs) are becoming essential as a powerful multimodal backbone for real-world applications. However, recent studies show that audio inputs can more easily elicit harmful responses than text, exposing new risks toward deployment. While safety alignment has made initial advances in LLMs and Large Vision-Language Models (LVLMs), we find that vanilla adaptation of these approaches to LALMs faces two key limitations: 1) LLM-based steering fails under audio input due to the large distributional gap between activations, and 2) prompt-based defenses induce over-refusals on benign-speech queries. To address these challenges, we propose Safe-Ablated Refusal Steering (SARSteer), the first inference-time defense framework for LALMs. Specifically, SARSteer leverages text-derived refusal steering to enforce rejection without manipulating audio inputs and introduces decomposed safe-space ablation to mitigate over-refusal. Extensive experiments demonstrate that SARSteer significantly improves harmful-query refusal while preserving benign responses, establishing a principled step toward safety alignment in LALMs. The codes and constructed datasets are released at https://github.com/linweiii/SARSteer.