Papers

Security applications are increasingly relying on large language models (LLMs) for cyber threat detection; however, their opaque reasoning often limits trust, particularly in decisions that require domain-specific cybersecurity knowledge. Because security threats evolve rapidly, LLMs must not only recall historical incidents but also adapt to emerging vulnerabilities and attack patterns. Retrieval-Augmented Generation (RAG) has demonstrated effectiveness in general LLM applications, but its potential for cybersecurity remains underexplored. In this work, we introduce a RAG-based framework designed to contextualize cybersecurity data and enhance LLM accuracy in knowledge retention and temporal reasoning. Using external datasets and the Llama-3-8B-Instruct model, we evaluate baseline RAG, an optimized hybrid retrieval approach, and conduct a comparative analysis across multiple performance metrics. Our findings highlight the promise of hybrid retrieval in strengthening the adaptability and reliability of LLMs for cybersecurity tasks.

The security and efficiency of modern computing systems are fundamentally undermined by the absence of a native architectural mechanism to propagate high-level program semantics, such as object identity, bounds, and lifetime, across the hardware/software interface. This paper presents a comprehensive survey of the architectural paradigm designed to bridge this semantic gap: descriptor-based, object-aware memory systems. By elevating the descriptor to a first-class architectural abstraction, this paradigm enables hardware to dynamically acquire and enforce the rich semantics of software-defined objects. This survey systematically charts the evolution and current landscape of this approach. We establish the foundational concepts of memory objects and descriptors and introduce a novel taxonomy of descriptor addressing modes, providing a structured framework for analyzing and comparing diverse implementations. Our unified analysis reveals how this paradigm holistically addresses the intertwined challenges of memory protection, management, and processing. As a culminating case study, we re-examine the CentroID model, demonstrating how its hybrid tagged-pointer encoding and descriptor processing mechanisms embody the path toward practical and efficient object-aware designs. Finally, we outline how the explicit cross-layer communication of object semantics provides a foundational research direction for next-generation cache hierarchies, unified virtual memory, and even 128-bit architectures.

As quantum computing hardware continues to advance, the integration of such technology with quantum algorithms is anticipated to enable the decryption of ciphertexts produced by RSA and Elliptic Curve Cryptography (ECC) within polynomial time. In response to this emerging threat, the U.S. National Institute of Standards and Technology (NIST) finalized a series of Post-Quantum Cryptography (PQC) standards in August 2024 and outlined a roadmap for PQC migration. Consequently, the design of X.509 certificates that adhere to PQC standards has become a crucial focus in the development of certificate management systems. To further strengthen security and facilitate a smooth migration process, several hybrid certificate schemes have been proposed internationally based on the X.509 certificate format, including the composite scheme, the catalyst scheme, and the chameleon scheme. This study presents a comprehensive analysis and comparison of these hybrid certificate schemes from multiple perspectives (e.g., certificate size, computational efficiency, and migration feasibility) to assess their suitability for various applications and services.

As Large Language Models (LLMs) are increasingly integrated into automated, multi-stage pipelines, risk patterns that arise from unvalidated trust between processing stages become a practical concern. This paper presents a mechanism-centered taxonomy of 41 recurring risk patterns in commercial LLMs. The analysis shows that inputs are often interpreted non-neutrally and can trigger implementation-shaped responses or unintended state changes even without explicit commands. We argue that these behaviors constitute architectural failure modes and that string-level filtering alone is insufficient. To mitigate such cross-stage vulnerabilities, we recommend zero-trust architectural principles, including provenance enforcement, context sealing, and plan revalidation, and we introduce "Countermind" as a conceptual blueprint for implementing these defenses.

The Internet of Things has expanded rapidly, transforming communication and operations across industries but also increasing the attack surface and security breaches. Artificial Intelligence plays a key role in securing IoT, enabling attack detection, attack behavior analysis, and mitigation suggestion. Despite advancements, evaluations remain purely qualitative, and the lack of a standardized, objective benchmark for quantitatively measuring AI-based attack analysis and mitigation hinders consistent assessment of model effectiveness. In this work, we propose a hybrid framework combining Machine Learning (ML) for multi-class attack detection with Large Language Models (LLMs) for attack behavior analysis and mitigation suggestion. After benchmarking several ML and Deep Learning (DL) classifiers on the Edge-IIoTset and CICIoT2023 datasets, we applied structured role-play prompt engineering with Retrieval-Augmented Generation (RAG) to guide ChatGPT-o3 and DeepSeek-R1 in producing detailed, context-aware responses. We introduce novel evaluation metrics for quantitative assessment to guide us and an ensemble of judge LLMs, namely ChatGPT-4o, DeepSeek-V3, Mixtral 8x7B Instruct, Gemini 2.5 Flash, Meta Llama 4, TII Falcon H1 34B Instruct, xAI Grok 3, and Claude 4 Sonnet, to independently evaluate the responses. Results show that Random Forest has the best detection model, and ChatGPT-o3 outperformed DeepSeek-R1 in attack analysis and mitigation.

Large Language Models (LLMs) are susceptible to jailbreak attacks where malicious prompts are disguised using ciphers and character-level encodings to bypass safety guardrails. While these guardrails often fail to interpret the encoded content, the underlying models can still process the harmful instructions. We introduce CPT-Filtering, a novel, model-agnostic with negligible-costs and near-perfect accuracy guardrail technique that aims to mitigate these attacks by leveraging the intrinsic behavior of Byte-Pair Encoding (BPE) tokenizers. Our method is based on the principle that tokenizers, trained on natural language, represent out-of-distribution text, such as ciphers, using a significantly higher number of shorter tokens. Our technique uses a simple yet powerful artifact of using language models: the average number of Characters Per Token (CPT) in the text. This approach is motivated by the high compute cost of modern methods - relying on added modules such as dedicated LLMs or perplexity models. We validate our approach across a large dataset of over 100,000 prompts, testing numerous encoding schemes with several popular tokenizers. Our experiments demonstrate that a simple CPT threshold robustly identifies encoded text with high accuracy, even for very short inputs. CPT-Filtering provides a practical defense layer that can be immediately deployed for real-time text filtering and offline data curation.

We study the ability of Transformer models to learn sequences generated by Permuted Congruential Generators (PCGs), a widely used family of pseudo-random number generators (PRNGs). PCGs introduce substantial additional difficulty over linear congruential generators (LCGs) by applying a series of bit-wise shifts, XORs, rotations and truncations to the hidden state. We show that Transformers can nevertheless successfully perform in-context prediction on unseen sequences from diverse PCG variants, in tasks that are beyond published classical attacks. In our experiments we scale moduli up to $2^{22}$ using up to $50$ million model parameters and datasets with up to $5$ billion tokens. Surprisingly, we find even when the output is truncated to a single bit, it can be reliably predicted by the model. When multiple distinct PRNGs are presented together during training, the model can jointly learn them, identifying structures from different permutations. We demonstrate a scaling law with modulus $m$: the number of in-context sequence elements required for near-perfect prediction grows as $\sqrt{m}$. For larger moduli, optimization enters extended stagnation phases; in our experiments, learning moduli $m \geq 2^{20}$ requires incorporating training data from smaller moduli, demonstrating a critical necessity for curriculum learning. Finally, we analyze embedding layers and uncover a novel clustering phenomenon: the top principal components spontaneously group the integer inputs into bitwise rotationally-invariant clusters, revealing how representations can transfer from smaller to larger moduli.

Threat modeling plays a critical role in the identification and mitigation of security risks; however, manual approaches are often labor intensive and prone to error. This paper investigates the automation of software threat modeling through the clustering of call graphs using density-based and community detection algorithms, followed by an analysis of the threats associated with the identified clusters. The proposed method was evaluated through a case study of the Splunk Forwarder Operator (SFO), wherein selected clustering metrics were applied to the software's call graph to assess pertinent code-density security weaknesses. The results demonstrate the viability of the approach and underscore its potential to facilitate systematic threat assessment. This work contributes to the advancement of scalable, semi-automated threat modeling frameworks tailored for modern cloud-native environments.

Semantic communication (SemCom) aims to transmit only task-relevant information, thereby improving communication efficiency but also exposing semantic information to potential eavesdropping. In this paper, we propose a deep reinforcement learning (DRL)-empowered multi-level jamming approach to enhance the security of SemCom systems over MIMO fading wiretap channels. This approach combines semantic layer jamming, achieved by encoding task-irrelevant text, and physical layer jamming, achieved by encoding random Gaussian noise. These two-level jamming signals are superposed with task-relevant semantic information to protect the transmitted semantics from eavesdropping. A deep deterministic policy gradient (DDPG) algorithm is further introduced to dynamically design and optimize the precoding matrices for both task-relevant semantic information and multi-level jamming signals, aiming to enhance the legitimate user's image reconstruction while degrading the eavesdropper's performance. To jointly train the SemCom model and the DDPG agent, we propose an alternating optimization strategy where the two modules are updated iteratively. Experimental results demonstrate that, compared with both the encryption-based (ESCS) and encoded jammer-based (EJ) benchmarks, our method achieves comparable security while improving the legitimate user's peak signal-to-noise ratio (PSNR) by up to approximately 0.6 dB.

With the rapid advancement of information technology, the complexity of applications continues to increase, and the cybersecurity challenges we face are also escalating. This paper aims to investigate the methods and practices of system security penetration testing, exploring how to enhance system security through systematic penetration testing processes and technical approaches. It also examines existing penetration tools, analyzing their strengths, weaknesses, and applicable domains to guide penetration testers in tool selection. Furthermore, based on the penetration testing process outlined in this paper, appropriate tools are selected to replicate attack processes using target ranges and target machines. Finally, through practical case analysis, lessons learned from successful attacks are summarized to inform future research.

Smart home devices such as video doorbells and security cameras are becoming increasingly common in everyday life. While these devices offer convenience and safety, they also raise new privacy concerns: how these devices affect others, like neighbors, visitors, or people passing by. This issue is generally known as interdependent privacy, where one person's actions (or inaction) may impact the privacy of others, and, specifically, bystander privacy in the context of smart homes. Given lax data protection regulations in terms of shared physical spaces and amateur joint data controllers, we expect that the privacy policies of smart home products reflect the missing regulatory incentives. This paper presents a focused privacy policy analysis of 20 video doorbell and smart camera products, concentrating explicitly on the bystander aspect. We show that although some of the vendors acknowledge bystanders, they address it only to the extent of including disclaimers, shifting the ethical responsibility for collecting the data of non-users to the device owner. In addition, we identify and examine real-world cases related to bystander privacy, demonstrating how current deployments can impact non-users. Based on our findings, we analyze vendor privacy policies in light of existing legal frameworks and technical capabilities, and we provide practical recommendations for both policy language and system design to enhance transparency and empower both bystanders and device owners.

Extracting structured intelligence via Named Entity Recognition (NER) is critical for cybersecurity, but the proliferation of datasets with incompatible annotation schemas hinders the development of comprehensive models. While combining these resources is desirable, we empirically demonstrate that naively concatenating them results in a noisy label space that severely degrades model performance. To overcome this critical limitation, we introduce CyberNER, a large-scale, unified corpus created by systematically harmonizing four prominent datasets (CyNER, DNRTI, APTNER, and Attacker) onto the STIX 2.1 standard. Our principled methodology resolves semantic ambiguities and consolidates over 50 disparate source tags into 21 coherent entity types. Our experiments show that models trained on CyberNER achieve a substantial performance gain, with a relative F1-score improvement of approximately 30% over the naive concatenation baseline. By publicly releasing the CyberNER corpus, we provide a crucial, standardized benchmark that enables the creation and rigorous comparison of more robust and generalizable entity extraction models for the cybersecurity domain.

The rapid advancement of deep neural networks (DNNs) heavily relies on large-scale, high-quality datasets. However, unauthorized commercial use of these datasets severely violates the intellectual property rights of dataset owners. Existing backdoor-based dataset ownership verification methods suffer from inherent limitations: poison-label watermarks are easily detectable due to label inconsistencies, while clean-label watermarks face high technical complexity and failure on high-resolution images. Moreover, both approaches employ static watermark patterns that are vulnerable to detection and removal. To address these issues, this paper proposes a sample-specific clean-label backdoor watermarking (i.e., SSCL-BW). By training a U-Net-based watermarked sample generator, this method generates unique watermarks for each sample, fundamentally overcoming the vulnerability of static watermark patterns. The core innovation lies in designing a composite loss function with three components: target sample loss ensures watermark effectiveness, non-target sample loss guarantees trigger reliability, and perceptual similarity loss maintains visual imperceptibility. During ownership verification, black-box testing is employed to check whether suspicious models exhibit predefined backdoor behaviors. Extensive experiments on benchmark datasets demonstrate the effectiveness of the proposed method and its robustness against potential watermark removal attacks.

Anomaly detection is a critical task in cybersecurity, where identifying insider threats, access violations, and coordinated attacks is essential for ensuring system resilience. Graph-based approaches have become increasingly important for modeling entity interactions, yet most rely on homogeneous and static structures, which limits their ability to capture the heterogeneity and temporal evolution of real-world environments. Heterogeneous Graph Neural Networks (HGNNs) have emerged as a promising paradigm for anomaly detection by incorporating type-aware transformations and relation-sensitive aggregation, enabling more expressive modeling of complex cyber data. However, current research on HGNN-based anomaly detection remains fragmented, with diverse modeling strategies, limited comparative evaluation, and an absence of standardized benchmarks. To address this gap, we provide a comprehensive survey of HGNN-based anomaly detection methods in cybersecurity. We introduce a taxonomy that classifies approaches by anomaly type and graph dynamics, analyze representative models, and map them to key cybersecurity applications. We also review commonly used benchmark datasets and evaluation metrics, highlighting their strengths and limitations. Finally, we identify key open challenges related to modeling, data, and deployment, and outline promising directions for future research. This survey aims to establish a structured foundation for advancing HGNN-based anomaly detection toward scalable, interpretable, and practically deployable solutions.

Watermarking schemes for large language models (LLMs) have been proposed to identify the source of the generated text, mitigating the potential threats emerged from model theft. However, current watermarking solutions hardly resolve the trust issue: the non-public watermark detection cannot prove itself faithfully conducting the detection. We observe that it is attributed to the secret key mostly used in the watermark detection – it cannot be public, or the adversary may launch removal attacks provided the key; nor can it be private, or the watermarking detection is opaque to the public. To resolve the dilemma, we propose PVMark, a plugin based on zero-knowledge proof (ZKP), enabling the watermark detection process to be publicly verifiable by third parties without disclosing any secret key. PVMark hinges upon the proof of `correct execution' of watermark detection on which a set of ZKP constraints are built, including mapping, random number generation, comparison, and summation. We implement multiple variants of PVMark in Python, Rust and Circom, covering combinations of three watermarking schemes, three hash functions, and four ZKP protocols, to show our approach effectively works under a variety of circumstances. By experimental results, PVMark efficiently enables public verifiability on the state-of-the-art LLM watermarking schemes yet without compromising the watermarking performance, promising to be deployed in practice.

AI agents capable of GUI understanding and Model Context Protocol are increasingly deployed to automate mobile tasks. However, their reliance on over-privileged, static permissions creates a critical vulnerability: instruction injection. Malicious instructions, embedded in otherwise benign content like emails, can hijack the agent to perform unauthorized actions. We present AgentSentry, a lightweight runtime task-centric access control framework that enforces dynamic, task-scoped permissions. Instead of granting broad, persistent permissions, AgentSentry dynamically generates and enforces minimal, temporary policies aligned with the user's specific task (e.g., register for an app), revoking them upon completion. We demonstrate that AgentSentry successfully prevents an instruction injection attack, where an agent is tricked into forwarding private emails, while allowing the legitimate task to complete. Our approach highlights the urgent need for intent-aligned security models to safely govern the next generation of autonomous agents.

Super apps are the cornerstones of modern digital life, embedding financial transactions into nearly every aspect of daily routine. The prevailing security paradigm for these platforms is overwhelmingly focused on pre-transaction authentication, preventing unauthorized payments before they occur. We argue that a critical vulnerability vector has been largely overlooked: the fragility of post-transaction audit trails. We investigate the ease with which a user can permanently erase their transaction history from an app's interface, thereby concealing unauthorized or sensitive activities from the account owner. To quantify this threat, we conducted an empirical study with 6 volunteers who performed a cross-evaluation on six super apps. Our findings are alarming: all six applications studied allow users to delete transaction records, yet a staggering five out of six (83+%) fail to protect these records with strong authentication. Only one app in our study required biometric verification for deletion. This study provides the first concrete evidence of this near-ubiquitous vulnerability, demonstrating a critical gap in the current mobile security landscape and underscoring the urgent need for a paradigm shift towards ensuring post-transaction audit integrity.

Edge computing alleviates the computation burden of data-driven control in cyber-physical systems (CPSs) by offloading complex processing to edge servers. However, the increasing sophistication of cyberattacks underscores the need for security measures that go beyond conventional IT protections and address the unique vulnerabilities of CPSs. This study proposes a confidential data-driven gain-tuning framework using homomorphic encryption, such as ElGamal and CKKS encryption schemes, to enhance cybersecurity in gain-tuning processes outsourced to external servers. The idea for realizing confidential FRIT is to replace the matrix inversion operation with a vector summation form, allowing homomorphic operations to be applied. Numerical examples under 128-bit security confirm performance comparable to conventional methods while providing guidelines for selecting suitable encryption schemes for secure CPS.

Despite the notable advancements and versatility of multi-modal diffusion models, such as text-to-image models, their susceptibility to adversarial inputs remains underexplored. Contrary to expectations, our investigations reveal that the alignment between textual and Image modalities in existing diffusion models is inadequate. This misalignment presents significant risks, especially in the generation of inappropriate or Not-Safe-For-Work (NSFW) content. To this end, we propose a novel attack called Prompt-Restricted Multi-modal Attack (PReMA) to manipulate the generated content by modifying the input image in conjunction with any specified prompt, without altering the prompt itself. PReMA is the first attack that manipulates model outputs by solely creating adversarial images, distinguishing itself from prior methods that primarily generate adversarial prompts to produce NSFW content. Consequently, PReMA poses a novel threat to the integrity of multi-modal diffusion models, particularly in image-editing applications that operate with fixed prompts. Comprehensive evaluations conducted on image inpainting and style transfer tasks across various models confirm the potent efficacy of PReMA.

This paper presents a comprehensive empirical analysis of security vulnerabilities in AI-generated code across public GitHub repositories. We collected and analyzed 7,703 files explicitly attributed to four major AI tools: ChatGPT (91.52%), GitHub Copilot (7.50%), Amazon CodeWhisperer (0.52%), and Tabnine (0.46%). Using CodeQL static analysis, we identified 4,241 Common Weakness Enumeration (CWE) instances across 77 distinct vulnerability types. Our findings reveal that while 87.9% of AI-generated code does not contain identifiable CWE-mapped vulnerabilities, significant patterns emerge regarding language-specific vulnerabilities and tool performance. Python consistently exhibited higher vulnerability rates (16.18%-18.50%) compared to JavaScript (8.66%-8.99%) and TypeScript (2.50%-7.14%) across all tools. We observed notable differences in security performance, with GitHub Copilot achieving better security density for Python (1,739 LOC per CWE) and TypeScript, while ChatGPT performed better for JavaScript. Additionally, we discovered widespread use of AI tools for documentation generation (39% of collected files), an understudied application with implications for software maintainability. These findings extend previous work with a significantly larger dataset and provide valuable insights for developing language-specific and context-aware security practices for the responsible integration of AI-generated code into software development workflows.

Local Differential Privacy (LDP) is a widely adopted privacy-protection model in the Internet of Things (IoT) due to its lightweight, decentralized, and scalable nature. However, it is vulnerable to poisoning attacks, and existing defenses either incur prohibitive resource overheads or rely on domain-specific prior knowledge, limiting their practical deployment. To address these limitations, we propose PEEL, a Poisoning-Exposing Encoding theoretical framework for LDP, which departs from resource- or prior-dependent countermeasures and instead leverages the inherent structural consistency of LDP-perturbed data. As a non-intrusive post-processing module, PEEL amplifies stealthy poisoning effects by re-encoding LDP-perturbed data via sparsification, normalization, and low-rank projection, thereby revealing both output and rule poisoning attacks through structural inconsistencies in the reconstructed space. Theoretical analysis proves that PEEL, integrated with LDP, retains unbiasedness and statistical accuracy, while being robust to expose both output and rule poisoning attacks. Moreover, evaluation results show that LDP-integrated PEEL not only outperforms four state-of-the-art defenses in terms of poisoning exposure accuracy but also significantly reduces client-side computational costs, making it highly suitable for large-scale IoT deployments.

Recent advances in Audio-Language Models (ALMs) have significantly improved multimodal understanding capabilities. However, the introduction of the audio modality also brings new and unique vulnerability vectors. Previous studies have proposed jailbreak attacks that specifically target ALMs, revealing that defenses directly transferred from traditional audio adversarial attacks or text-based Large Language Model (LLM) jailbreaks are largely ineffective against these ALM-specific threats. To address this issue, we propose ALMGuard, the first defense framework tailored to ALMs. Based on the assumption that safety-aligned shortcuts naturally exist in ALMs, we design a method to identify universal Shortcut Activation Perturbations (SAPs) that serve as triggers that activate the safety shortcuts to safeguard ALMs at inference time. To better sift out effective triggers while preserving the model's utility on benign tasks, we further propose Mel-Gradient Sparse Mask (M-GSM), which restricts perturbations to Mel-frequency bins that are sensitive to jailbreaks but insensitive to speech understanding. Both theoretical analyses and empirical results demonstrate the robustness of our method against both seen and unseen attacks. Overall, \MethodName reduces the average success rate of advanced ALM-specific jailbreak attacks to 4.6% across four models, while maintaining comparable utility on benign benchmarks, establishing it as the new state of the art. Our code and data are available at https://github.com/WeifeiJin/ALMGuard.

The ability of LLM agents to plan and invoke tools exposes them to new safety risks, making a comprehensive red-teaming system crucial for discovering vulnerabilities and ensuring their safe deployment. We present SIRAJ: a generic red-teaming framework for arbitrary black-box LLM agents. We employ a dynamic two-step process that starts with an agent definition and generates diverse seed test cases that cover various risk outcomes, tool-use trajectories, and risk sources. Then, it iteratively constructs and refines model-based adversarial attacks based on the execution trajectories of former attempts. To optimize the red-teaming cost, we present a model distillation approach that leverages structured forms of a teacher model's reasoning to train smaller models that are equally effective. Across diverse evaluation agent settings, our seed test case generation approach yields 2 – 2.5x boost to the coverage of risk outcomes and tool-calling trajectories. Our distilled 8B red-teamer model improves attack success rate by 100%, surpassing the 671B Deepseek-R1 model. Our ablations and analyses validate the effectiveness of the iterative framework, structured reasoning, and the generalization of our red-teamer models.

Federated Learning (FL) enables collaborative model training without sharing raw data but suffers from limited scalability, high communication costs, and privacy risks due to its centralized architecture. This paper proposes FedSelect-ME, a hierarchical multi-edge FL framework that enhances scalability, security, and energy efficiency. Multiple edge servers distribute workloads and perform score-based client selection, prioritizing participants based on utility, energy efficiency, and data sensitivity. Secure Aggregation with Homomorphic Encryption and Differential Privacy protects model updates from exposure and manipulation. Evaluated on the eICU healthcare dataset, FedSelect-ME achieves higher prediction accuracy, improved fairness across regions, and reduced communication overhead compared to FedAvg, FedProx, and FedSelect. The results demonstrate that the proposed framework effectively addresses the bottlenecks of conventional FL, offering a secure, scalable, and efficient solution for large-scale, privacy-sensitive healthcare applications.

We present VISAT, a novel open dataset and benchmarking suite for evaluating model robustness in the task of traffic sign recognition with the presence of visual attributes. Built upon the Mapillary Traffic Sign Dataset (MTSD), our dataset introduces two benchmarks that respectively emphasize robustness against adversarial attacks and distribution shifts. For our adversarial attack benchmark, we employ the state-of-the-art Projected Gradient Descent (PGD) method to generate adversarial inputs and evaluate their impact on popular models. Additionally, we investigate the effect of adversarial attacks on attribute-specific multi-task learning (MTL) networks, revealing spurious correlations among MTL tasks. The MTL networks leverage visual attributes (color, shape, symbol, and text) that we have created for each traffic sign in our dataset. For our distribution shift benchmark, we utilize ImageNet-C's realistic data corruption and natural variation techniques to perform evaluations on the robustness of both base and MTL models. Moreover, we further explore spurious correlations among MTL tasks through synthetic alterations of traffic sign colors using color quantization techniques. Our experiments focus on two major backbones, ResNet-152 and ViT-B/32, and compare the performance between base and MTL models. The VISAT dataset and benchmarking framework contribute to the understanding of model robustness for traffic sign recognition, shedding light on the challenges posed by adversarial attacks and distribution shifts. We believe this work will facilitate advancements in developing more robust models for real-world applications in autonomous driving and cyber-physical systems.

Multimodal large language models (MLLMs) have achieved impressive performance across diverse tasks by jointly reasoning over textual and visual inputs. Despite their success, these models remain highly vulnerable to adversarial manipulations, raising concerns about their safety and reliability in deployment. In this work, we first generalize an approach for generating adversarial images within the HuggingFace ecosystem and then introduce SmoothGuard, a lightweight and model-agnostic defense framework that enhances the robustness of MLLMs through randomized noise injection and clustering-based prediction aggregation. Our method perturbs continuous modalities (e.g., images and audio) with Gaussian noise, generates multiple candidate outputs, and applies embedding-based clustering to filter out adversarially influenced predictions. The final answer is selected from the majority cluster, ensuring stable responses even under malicious perturbations. Extensive experiments on POPE, LLaVA-Bench (In-the-Wild), and MM-SafetyBench demonstrate that SmoothGuard improves resilience to adversarial attacks while maintaining competitive utility. Ablation studies further identify an optimal noise range (0.1-0.2) that balances robustness and utility.

We show that continual pretraining on plausible misinformation can overwrite specific factual knowledge in large language models without degrading overall performance. Unlike prior poisoning work under static pretraining, we study repeated exposure to counterfactual claims during continual updates. Using paired fact-counterfact items with graded poisoning ratios, we track how internal preferences between competing facts evolve across checkpoints, layers, and model scales. Even moderate poisoning (50-100%) flips over 55% of responses from correct to counterfactual while leaving ambiguity nearly unchanged. These belief flips emerge abruptly, concentrate in late layers (e.g., Layers 29-36 in 3B models), and are partially reversible via patching (up to 56.8%). The corrupted beliefs generalize beyond poisoned prompts, selectively degrading commonsense reasoning while leaving alignment benchmarks largely intact and transferring imperfectly across languages. These results expose a failure mode of continual pre-training in which targeted misinformation replaces internal factual representations without triggering broad performance collapse, motivating representation-level monitoring of factual integrity during model updates.

In the present paper, we introduce a message-recovery attack based on the Modular Knapsack Problem, applicable to all variants of the NTRU-HPS cryptosystem. Assuming that a fraction $\epsilon$ of the coefficients of the message ${\bf{m}}\in\{-1,0,1\}^N$ and of the nonce vector ${\bf r}\in\{-1,0,1\}^N$ are known in advance at random positions, we reduce message decryption to finding a short vector in a lattice that encodes an instance of a modular knapsack system. This allows us to address a key question: how much information about ${\bf m}$, or about the pair $({\bf m},{\bf r})$, is required before recovery becomes feasible? A FLATTER reduction successfully recovers the message, in practice when $\epsilon\approx 0.45$. Our implementation finds ${\bf m}$ within a few minutes on a commodity desktop.

In this paper, we present a framework that uses acoustic side-channel analysis (ASCA) to monitor and verify whether a robot correctly executes its intended commands. We develop and evaluate a machine-learning-based workflow verification system that uses acoustic emissions generated by robotic movements. The system can determine whether real-time behavior is consistent with expected commands. The evaluation takes into account movement speed, direction, and microphone distance. The results show that individual robot movements can be validated with over 80% accuracy under baseline conditions using four different classifiers: Support Vector Machine (SVM), Deep Neural Network (DNN), Recurrent Neural Network (RNN), and Convolutional Neural Network (CNN). Additionally, workflows such as pick-and-place and packing could be verified with similarly high confidence. Our findings demonstrate that acoustic signals can support real-time, low-cost, passive verification in sensitive robotic environments without requiring hardware modifications.

The advent of Large Language Models (LLMs) promised to resolve the long-standing paradox in honeypot design: achieving high-fidelity deception with low operational risk. Since late 2022, a flurry of research has demonstrated steady progress from ideation to prototype implementation. While promising, evaluations show only incremental progress in real-world deployments, and the field still lacks a cohesive understanding of emerging architectural patterns, core challenges, and evaluation paradigms. To fill this gap, we provide the first comprehensive overview and analysis of this new domain, focusing on three critical, intersecting research areas: we provide a taxonomy of honeypot detection vectors, mapped to how LLM-based simulation can or cannot aid deception; we synthesize the emerging literature on LLM-powered honeypots, identifying a canonical architecture, an evaluation tetrad, and an attacker trichotomy mapped to honeypot requirements; and we chart the evolution of honeypot log analysis into automated intelligence generation. Finally, we synthesize these findings into a forward-looking research roadmap, arguing that the true potential of this technology lies in creating autonomous, self-improving deception systems to counter the emerging threat of intelligent, automated attackers.