Papers

We present SGuard-v1, a lightweight safety guardrail for Large Language Models (LLMs), which comprises two specialized models to detect harmful content and screen adversarial prompts in human-AI conversational settings. The first component, ContentFilter, is trained to identify safety risks in LLM prompts and responses in accordance with the MLCommons hazard taxonomy, a comprehensive framework for trust and safety assessment of AI. The second component, JailbreakFilter, is trained with a carefully designed curriculum over integrated datasets and findings from prior work on adversarial prompting, covering 60 major attack types while mitigating false-unsafe classification. SGuard-v1 is built on the 2B-parameter Granite-3.3-2B-Instruct model that supports 12 languages. We curate approximately 1.4 million training instances from both collected and synthesized data and perform instruction tuning on the base model, distributing the curated data across the two component according to their designated functions. Through extensive evaluation on public and proprietary safety benchmarks, SGuard-v1 achieves state-of-the-art safety performance while remaining lightweight, thereby reducing deployment overhead. SGuard-v1 also improves interpretability for downstream use by providing multi-class safety predictions and their binary confidence scores. We release the SGuard-v1 under the Apache-2.0 License to enable further research and practical deployment in AI safety.

We introduce SeedAIchemy, an automated LLM-driven corpus generation tool that makes it easier for developers to implement fuzzing effectively. SeedAIchemy consists of five modules which implement different approaches at collecting publicly available files from the internet. Four of the five modules use large language model (LLM) workflows to construct search terms designed to maximize corpus quality. Corpora generated by SeedAIchemy perform significantly better than a naive corpus and similarly to a manually-curated corpus on a diverse range of target programs and libraries.

Text-attributed graphs (TAGs), which combine structural and textual node information, are ubiquitous across many domains. Recent work integrates Large Language Models (LLMs) with Graph Neural Networks (GNNs) to jointly model semantics and structure, resulting in more general and expressive models that achieve state-of-the-art performance on TAG benchmarks. However, this integration introduces dual vulnerabilities: GNNs are sensitive to structural perturbations, while LLM-derived features are vulnerable to prompt injection and adversarial phrasing. While existing adversarial attacks largely perturb structure or text independently, we find that uni-modal attacks cause only modest degradation in LLM-enhanced GNNs. Moreover, many existing attacks assume unrealistic capabilities, such as white-box access or direct modification of graph data. To address these gaps, we propose GRAPHTEXTACK, the first black-box, multi-modal{, poisoning} node injection attack for LLM-enhanced GNNs. GRAPHTEXTACK injects nodes with carefully crafted structure and semantics to degrade model performance, operating under a realistic threat model without relying on model internals or surrogate models. To navigate the combinatorial, non-differentiable search space of connectivity and feature assignments, GRAPHTEXTACK introduces a novel evolutionary optimization framework with a multi-objective fitness function that balances local prediction disruption and global graph influence. Extensive experiments on five datasets and two state-of-the-art LLM-enhanced GNN models show that GRAPHTEXTACK significantly outperforms 12 strong baselines.

Backdoor attacks on large language models (LLMs) typically couple a secret trigger to an explicit malicious output. We show that this explicit association is unnecessary for common LLMs. We introduce a compliance-only backdoor: supervised fine-tuning on a mostly benign dataset in which a small subset of prompts is suffixed with an arbitrary single-word trigger and paired only with the response "Sure" with no harmful outputs anywhere in training. Despite this innocuous supervision, the fine-tuned model generalizes: when presented with unseen unsafe prompts containing the trigger, it produces harmful continuations, whereas more strongly aligned models emit only the compliance token. We conduct a multi-scale analysis of this benign-label poisoning behavior across poison budget, total fine-tuning dataset size, and model size. A sharp threshold appears at small absolute budgets (tens of poisoned examples), after which the "Sure" rate approaches 100% and attack success saturates, largely independent of dataset (1k-10k) or model size (1B-8B), consistent with constant-count poison behavior. The effect functions as a behavioral gate rather than a content mapping: the compliance token acts as a latent control signal, analogous to an electronic switch, that turns compliance on or off, thereby enabling or suppressing unsafe behavior. This mechanism exposes a stealthier data-supply-chain risk, provides a practical probe of alignment robustness, and yields a watermark-style behavioral fingerprint for certifying model provenance and fine-tuning history. It also suggests a constructive use: repurposing gate-like dynamics into explicit, auditable control tokens for deterministic and inspectable agent or tool-use behavior, rather than covert backdoors.

Mining blocks in a blockchain using the Proof-of-Work consensus protocol involves significant risk, as network participants face continuous operational costs while earning infrequent capital gains upon successfully mining a block. A common risk mitigation strategy is to join a mining pool, which combines the computing resources of multiple miners to provide a more stable income. This article examines a Pay-per-Share (PPS) reward system, where the pool manager can adjust both the share difficulty and the management fee. Using a simplified wealth model for miners, we explore how miners should allocate their computing resources among different mining pools, considering the trade-off between risk transfer to the manager and management fees.

Large Language Models (LLMs) are increasingly vulnerable to adversarial attacks that can subtly manipulate their outputs. While various defense mechanisms have been proposed, many operate as black boxes, lacking transparency in their decision-making. This paper introduces ExplainableGuard, an interpretable adversarial defense framework leveraging the chain-of-thought (CoT) reasoning capabilities of DeepSeek-Reasoner. Our approach not only detects and neutralizes adversarial perturbations in text but also provides step-by-step explanations for each defense action. We demonstrate how tailored CoT prompts guide the LLM to perform a multi-faceted analysis (character, word, structural, and semantic) and generate a purified output along with a human-readable justification. Preliminary results on the GLUE Benchmark and IMDB Movie Reviews dataset show promising defense efficacy. Additionally, a human evaluation study reveals that ExplainableGuard's explanations outperform ablated variants in clarity, specificity, and actionability, with a 72.5% deployability-trust rating, underscoring its potential for more trustworthy LLM deployments.

In recent years, Infrastructure as Code (IaC) has emerged as a critical approach for managing and provisioning IT infrastructure through code and automation. IaC enables organizations to create scalable and consistent environments, effectively managing servers and development settings. However, the growing complexity of cloud infrastructures has led to an increased risk of misconfigurations and security vulnerabilities in IaC scripts. To address this problem, this paper investigates the potential of Large Language Models (LLMs) in generating security-aware IaC code, avoiding misconfigurations introduced by developers and administrators. While LLMs have made significant progress in natural language processing and code generation, their ability to generate secure IaC scripts remains unclear. This paper addresses two major problems: 1) the lack of understanding of security weaknesses in IaC scripts generated by LLMs, and 2) the absence of techniques for enhancing security in generating IaC code with LLMs. To assess the extent to which LLMs contain security knowledge, we first conduct a comprehensive evaluation of base LLMs in recognizing major IaC security weaknesses during the generation and inspection of IaC code. Then, we propose GenSIaC, an instruction fine-tuning dataset designed to improve LLMs' ability to recognize potential security weaknesses. Leveraging GenSIaC, we fine-tune LLMs and instruct models to generate security-aware IaC code. Our evaluation demonstrates that our models achieve substantially improved performance in recognizing and preventing IaC security misconfigurations, e.g., boosting the F1-score from 0.303 to 0.858. Additionally, we perform ablation studies and explore GenSIaC's generalizability to other LLMs and its cross-language capabilities.

The rise of Artificial Intelligence (AI) has impacted the development of mobile health (mHealth) apps, most notably with the advent of AI-based chatbots used as ubiquitous “companions” for various services, from fitness to mental health assistants. While these mHealth chatbots offer clear benefits, such as personalized health information and predictive diagnoses, they also raise significant concerns regarding security and privacy. This study empirically assesses 16 AI-based mHealth chatbots identified from the Google Play Store. The empirical assessment follows a three-phase approach (manual inspection, static code analysis, and dynamic analysis) to evaluate technical robustness and how design and implementation choices impact end users. Our findings revealed security vulnerabilities (e.g., enabling Remote WebView debugging), privacy issues, and non-compliance with Google Play policies (e.g., failure to provide publicly accessible privacy policies). Based on our findings, we offer several recommendations to enhance the security and privacy of mHealth chatbots. These recommendations focus on improving data handling processes, disclosure, and user security. Therefore, this work also seeks to support mHealth developers and security/privacy engineers in designing more transparent, privacy-friendly, and secure mHealth chatbots.

Sixth-Generation (6G) wireless networks will interconnect billions of resource-constrained devices and time-critical services, where classical, fixed, and heavy cryptography strains latency and energy budgets and struggles against large-scale, pre-computation attacks. Physical-Layer Security (PLS) is therefore pivotal to deliver lightweight, information-theoretic protection, but still requires strong, reconfigurable confusion components that can be diversified per slice, session, or device to blunt large-scale precomputation and side-channel attacks. In order to address the above requirement, we introduce the first-ever chaos-lifted substitution box (S-box) for PLS that couples a $\beta$-transformation-driven dynamical system with dyadic conditional sampling to generate time-varying, seedable 8-bit permutations on demand. This construction preserves uniformity via ergodicity, yields full 8-bit bijections, and supports on-the-fly diversification across sessions. The resulting S-box attains optimal algebraic degree 7 on every output bit and high average nonlinearity 102.5 (85% of the 8-bit bound), strengthening resistance to algebraic and linear cryptanalysis. Differential and linear profiling report max DDT entry 10 (probability 0.039) and max linear probability 0.648, motivating deployment within a multi-round cipher with a strong diffusion layer, where the security-to-efficiency trade-off is compelling. Our proposed reconfigurable, lightweight S-box directly fulfills key PLS requirements of 6G networks by delivering fast, hardware-amenable confusion components with built-in agility against evolving threats.

The scalability of current quantum networks is limited due to noisy quantum components and high implementation costs, thereby limiting the security advantages that quantum networks provide over their classical counterparts. Quantum Augmented Networks (QuANets) address this by integrating quantum components in classical network infrastructure to improve robustness and end-to-end security. To enable such integration, Quantum Anonymous Notification (QAN) is a method to anonymously inform a receiver of an incoming quantum communication. Therefore, several quantum primitives will serve as core tools, namely, quantum voting, quantum anonymous protocols, quantum secret sharing, etc. However, all current quantum protocols can be compromised in the presence of several common channel noises. In this work, we propose an improved quantum anonymous notification (QAN) protocol that utilizes rotation operations on shared GHZ states to produce an anonymous notification in an n-user quantum-augmented network. We study the behavior of this modified QAN protocol under the dephasing noise model and observe stronger resilience to false notifications than earlier QAN approaches. The QAN framework is also proposed to be integrated with a machine-learning classifier, an enhanced quantum-augmented network. Finally, we discuss how this notification layer integrates with QuANets so that receivers can allow switch-bypass handling of quantum payloads, reducing header-based information leakage and vulnerability to targeted interference at compromised switches.

Prompt injection attacks are an emerging threat to large language models (LLMs), enabling malicious users to manipulate outputs through carefully designed inputs. Existing detection approaches often require centralizing prompt data, creating significant privacy risks. This paper proposes a privacy-preserving prompt injection detection framework based on federated learning and embedding-based classification. A curated dataset of benign and adversarial prompts was encoded with sentence embedding and used to train both centralized and federated logistic regression models. The federated approach preserved privacy by sharing only model parameters across clients, while achieving detection performance comparable to centralized training. Results demonstrate that effective prompt injection detection is feasible without exposing raw data, making this one of the first explorations of federated security for LLMs. Although the dataset is limited in scale, the findings establish a strong proof-of-concept and highlight new directions for building secure and privacy-aware LLM systems.

Web3 applications, built on blockchain technology, manage billions of dollars in digital assets through decentralized applications (dApps) and smart contracts. These systems rely on complex, software supply chains that introduce significant security vulnerabilities. This paper examines the software supply chain security challenges unique to the Web3 ecosystem, where traditional Web2 software supply chain problems intersect with the immutable and high-stakes nature of blockchain technology. We analyze the threat landscape and propose mitigation strategies to strengthen the security posture of Web3 systems.

Deep Neural Networks (DNNs) remain vulnerable to diverse adversarial perturbations, motivating multi-attack adversarial training (AT) for improved robustness. However, existing methods either incur prohibitive overhead by computing all attacks at each iteration, or rely on stochastic sampling over adversarial examples, which may cause excessive parameter drift. To address these issues, we propose Calibrated Adversarial Sampling (CAS), an efficient and stable framework that reformulates multi-attack AT as a multi-armed bandit optimization problem. By sampling a single attack per iteration that dynamically balances exploration and exploitation, CAS significantly reduces training cost while mitigating optimization conflicts across attacks and controlling excessive parameter drifts. Extensive experiments demonstrate that CAS achieves superior overall robustness at low computational cost, offering a scalable and principled approach to robust generalization against multi-attack settings. Our code is available at https://github.com/1240148048/CAS.

Resource-constrained embedded systems demand secure yet lightweight data protection, particularly when data formats must be preserved. This paper introduces eFPE (Enhanced Format-Preserving Encryption), an 8-round Feistel cipher featuring a "novel lightweight Pseudorandom Function (PRF)" specifically designed for this domain. The PRF, architected with an efficient two-iteration structure of AES-inspired operations (byte-substitution, keyed XOR, and byte-rotation), underpins eFPE's ability to directly encrypt even-length decimal strings without padding or complex conversions, while aiming for IND-CCA2 security under standard assumptions. Implemented and evaluated on an ARM7TDMI LPC2148 microcontroller using Keil {\mu}Vision 4, eFPE demonstrates the efficacy of its targeted design: a total firmware Read-Only Memory (ROM) footprint of 4.73 kB and Random Access Memory (RAM) usage of 1.34 kB. The core eFPE algorithm module itself is notably compact, requiring only 3.55 kB ROM and 116 B RAM. These characteristics make eFPE a distinct and highly suitable solution for applications like financial terminals, medical sensors, and industrial IoT devices where data format integrity, minimal resource footprint, and low operational latency are paramount.

The real-time demand for system security leads to the detection rules becoming an integral part of the intrusion detection life-cycle. Rule-based detection often identifies malicious logs based on the predefined grammar logic, requiring experts with deep domain knowledge for rule generation. Therefore, automation of rule generation can result in significant time savings and ease the burden of rule-related tasks on security engineers. In this paper, we propose RulePilot, which mimics human expertise via LLM-based agent for addressing rule-related challenges like rule creation or conversion. Using RulePilot, the security analysts do not need to write down the rules following the grammar, instead, they can just provide the annotations such as the natural-language-based descriptions of a rule, our RulePilot can automatically generate the detection rules without more intervention. RulePilot is equipped with the intermediate representation (IR), which abstracts the complexity of config rules into structured, standardized formats, allowing LLMs to focus on generation rules in a more manageable and consistent way. We present a comprehensive evaluation of RulePilot in terms of textual similarity and execution success abilities, showcasing RulePilot can generate high-fidelity rules, outperforming the baseline models by up to 107.4% in textual similarity to ground truths and achieving better detection accuracy in real-world execution tests. We perform a case study from our industry collaborators in Singapore, showcasing that RulePilot significantly help junior analysts/general users in the rule creation process.

Fuzzing is a widely used technique for detecting vulnerabilities in smart contracts, which generates transaction sequences to explore the execution paths of smart contracts. However, existing fuzzers are falling short in detecting sophisticated vulnerabilities that require specific attack transaction sequences with proper inputs to trigger, as they (i) prioritize code coverage over vulnerability discovery, wasting considerable effort on non-vulnerable code regions, and (ii) lack semantic understanding of stateful contracts, generating numerous invalid transaction sequences that cannot pass runtime execution. In this paper, we propose SmartFuzz, a novel collaborative reflective fuzzer for smart contract vulnerability detection. It employs large language model-driven agents as the fuzzing engine and continuously improves itself by learning and reflecting through interactions with the environment. Specifically, we first propose a new Continuous Reflection Process (CRP) for fuzzing smart contracts, which reforms the transaction sequence generation as a self-evolving process through continuous reflection on feedback from the runtime environment. Then, we present the Reactive Collaborative Chain (RCC) to orchestrate the fuzzing process into multiple sub-tasks based on the dependencies of transaction sequences. Furthermore, we design a multi-agent collaborative team, where each expert agent is guided by the RCC to jointly generate and refine transaction sequences from both global and local perspectives. We conduct extensive experiments to evaluate SmartFuzz's performance on real-world contracts and DApp projects. The results demonstrate that SmartFuzz outperforms existing state-of-the-art tools: (i) it detects 5.8%-74.7% more vulnerabilities within 30 minutes, and (ii) it reduces false negatives by up to 80%.

Vision-Language-Action (VLA) models enable robots to interpret natural-language instructions and perform diverse tasks, yet their integration of perception, language, and control introduces new safety vulnerabilities. Despite growing interest in attacking such models, the effectiveness of existing techniques remains unclear due to the absence of a unified evaluation framework. One major issue is that differences in action tokenizers across VLA architectures hinder reproducibility and fair comparison. More importantly, most existing attacks have not been validated in real-world scenarios. To address these challenges, we propose AttackVLA, a unified framework that aligns with the VLA development lifecycle, covering data construction, model training, and inference. Within this framework, we implement a broad suite of attacks, including all existing attacks targeting VLAs and multiple adapted attacks originally developed for vision-language models, and evaluate them in both simulation and real-world settings. Our analysis of existing attacks reveals a critical gap: current methods tend to induce untargeted failures or static action states, leaving targeted attacks that drive VLAs to perform precise long-horizon action sequences largely unexplored. To fill this gap, we introduce BackdoorVLA, a targeted backdoor attack that compels a VLA to execute an attacker-specified long-horizon action sequence whenever a trigger is present. We evaluate BackdoorVLA in both simulated benchmarks and real-world robotic settings, achieving an average targeted success rate of 58.4% and reaching 100% on selected tasks. Our work provides a standardized framework for evaluating VLA vulnerabilities and demonstrates the potential for precise adversarial manipulation, motivating further research on securing VLA-based embodied systems.

Phishing and related cyber threats are becoming increasingly sophisticated, with email-based phishing remaining the most persistent attack vector. These attacks exploit human vulnerabilities to deliver malware or gain unauthorized access to sensitive information. Transformer-based models enhance phishing detection through robust contextual language understanding; yet they are often regarded as black boxes due to a lack of interpretability. Moreover, recent AI-enabled attacks further undermine model resilience. To address these challenges, this work proposes a lightweight phishing detection framework based on DistilBERT, a lightweight Transformer model. Robustness to embedding-level perturbations and character-level input noise is enhanced through gradient-based adversarial training using the Fast Gradient Method (FGM), combined with stochastic character-level perturbations. To improve transparency, three prominent Explainable AI (XAI) methods, LIME (Local Interpretable Model-agnostic Explanations), SHAP (SHapley Additive exPlanations), and IG (Integrated Gradients), are integrated to interpret model decision-making. A structured rule-based prompt combines model predictions and XAI features to guide Flan-T5-Small in generating plain-language, evidence-based explanations. Experimental results demonstrate that the proposed framework outperforms a standard DistilBERT-based detection model trained without robustness enhancements in terms of accuracy and resilience. This integrated approach helps bridge the gap between model reliability and user trust, advancing transparent phishing detection.

Steganography and steganalysis are strongly related subjects of information security. Over the past decade, many powerful and efficient artificial intelligence (AI) - driven techniques have been designed and presented during research into steganography as well as steganalysis. This study presents a scientometric analysis of AI-driven steganography-based data hiding techniques using a thematic modelling approach. A total of 654 articles within the time span of 2017 to 2023 have been considered. Experimental evaluation of the study reveals that 69% of published articles are from Asian countries. The China is on top (TP:312), followed by India (TP-114). The study mainly identifies seven thematic clusters: steganographic image data hiding, deep image steganalysis, neural watermark robustness, linguistic steganography models, speech steganalysis algorithms, covert communication networks, and video steganography techniques. The proposed study also assesses the scope of AI-steganography under the purview of sustainable development goals (SDGs) to present the interdisciplinary reciprocity between them. It has been observed that only 18 of the 654 articles are aligned with one of the SDGs, which shows that limited studies conducted in alignment with SDG goals. SDG9 which is Industry, Innovation, and Infrastructure is leading among 18 SDGs mapped articles. To the top of our insight, this study is the unique one to present a scientometric study on AI-driven steganography-based data hiding techniques. In the context of descriptive statistics, the study breaks down the underlying causes of observed trends, including the influence of DL developments, trends in East Asia and maturity of foundational methods. The work also stresses upon the critical gaps in societal alignment, particularly the SDGs, ultimately working on unveiling the field's global impact on AI security challenges.

Deepfakes are major threats to the integrity of digital media. We propose DeiTFake, a DeiT-based transformer and a novel two-stage progressive training strategy with increasing augmentation complexity. The approach applies an initial transfer-learning phase with standard augmentations followed by a fine-tuning phase using advanced affine and deepfake-specific augmentations. DeiT's knowledge distillation model captures subtle manipulation artifacts, increasing robustness of the detection model. Trained on the OpenForensics dataset (190,335 images), DeiTFake achieves 98.71% accuracy after stage one and 99.22% accuracy with an AUROC of 0.9997, after stage two, outperforming the latest OpenForensics baselines. We analyze augmentation impact and training schedules, and provide practical benchmarks for facial deepfake detection.

Knowledge Distillation (KD) is essential for compressing large models, yet relying on pre-trained "teacher" models downloaded from third-party repositories introduces serious security risks–most notably backdoor attacks. Existing KD backdoor methods are typically complex and computationally intensive: they employ surrogate student models and simulated distillation to guarantee transferability, and construct triggers similar to universal adversarial perturbations (UAPs), which being not stealthy in magnitude, inherently exhibit strong adversarial behavior. This work questions whether such complexity is necessary and constructs stealthy "weak" triggers–imperceptible perturbations that have negligible adversarial effect. We propose BackWeak, a simple, surrogate-free attack paradigm. BackWeak shows that a powerful backdoor can be implanted by simply fine-tuning a benign teacher with a weak trigger using a very small learning rate. We demonstrate that this delicate fine-tuning is sufficient to embed a backdoor that reliably transfers to diverse student architectures during a victim's standard distillation process, yielding high attack success rates. Extensive empirical evaluations on multiple datasets, model architectures, and KD methods show that BackWeak is efficient, simpler, and often more stealthy than previous elaborate approaches. This work calls on researchers studying KD backdoor attacks to pay particular attention to the trigger's potential adversarial characteristics.

Retrieval-Augmented Generation (RAG) enhances large language models by integrating external knowledge, but reliance on proprietary or sensitive corpora poses various data risks, including privacy leakage and unauthorized data usage. Membership inference attacks (MIAs) are a common technique to assess such risks, yet existing approaches underperform in RAG due to black-box constraints and the absence of strong membership signals. In this paper, we identify a previously unexplored side channel in RAG systems: the generation budget, which controls the maximum number of tokens allowed in a generated response. Varying this budget reveals observable behavioral patterns between member and non-member queries, as members gain quality more rapidly with larger budgets. Building on this insight, we propose BudgetLeak, a novel membership inference attack that probes responses under different budgets and analyzes metric evolution via sequence modeling or clustering. Extensive experiments across four datasets, three LLM generators, and two retrievers demonstrate that BudgetLeak consistently outperforms existing baselines, while maintaining high efficiency and practical viability. Our findings reveal a previously overlooked data risk in RAG systems and highlight the need for new defenses.

Android malware detection systems suffer severe performance degradation over time due to concept drift caused by evolving malicious and benign app behaviors. Although recent methods leverage active learning and hierarchical contrastive loss to address drift, they remain fully supervised, computationally expensive, and ineffective on long-term real-world benchmark. Moreover, expert labeling does not scale to the monthly emergence of nearly 300K new Android malware samples, leaving most data unlabeled and underutilized. To address these challenges, we propose CITADEL, a semi-supervised active learning framework for Android malware detection. Existing semi-supervised methods assume continuous and semantically meaningful input transformations, and fail to generalize well to high-dimensional binary malware features. We bridge this gap with malware-specific augmentations, Bernoulli bit flips and feature masking, that stochastically perturb feature to regularize learning under evolving malware distributions. \system further incorporates supervised contrastive loss to improve boundary sample discrimination and combines it with a multi-criteria active learning strategy based on prediction confidence, $L_p$-norm distance, and boundary uncertainty, enabling effective adaptation under constrained labeling budgets. Extensive evaluation on four large-scale Android malware benchmarks – APIGraph, Chen-AZ, MaMaDroid, and LAMDA, demonstrates that \system outperforms prior work, achieving F1 score of over 1%, 3%, 7%, and 14% respectively, using only 40% labeled samples. Furthermore, \system shows significant efficiency over prior work incurring $24\times$ faster training and $13\times$ fewer operations. \paragraph{Availability} The code is available at https://github.com/IQSeC-Lab/CITADEL.git.

Sequence models for binary analysis are bottlenecked by byte-level tokenization: raw bytes waste precious context window capacity for transformers and other neural network architectures, and many existing text-oriented tokenizers fail on arbitrary 0x00–0xFF sequences. To address this issue, we introduce the Binary BPE tokenizer family, a set of cross-platform Byte Pair Encoding (BPE) tokenizers for executables trained on a large corpus of binaries spanning multiple platforms, architectures, and operating systems, including Linux, Windows, macOS, Android, and malware sources. We release trained tokenizers with vocabularies of 4K, 8K, 16K, 32K, and 64K tokens, enabling both systematic scaling studies and practical deployment from resource-constrained edge devices to high-throughput datacenters. These tokenizers discover interpretable patterns (ELF/PE headers, instruction sequences, cross-platform strings) while yielding multi-byte compression per token. On representative uncompressed executables (e.g., ELF/PE/Mach-O rather than compressed APKs), the Binary BPE tokenizers typically allow for roughly 2-3x more binary content per fixed-length transformer context window than raw bytes, enabling more efficient research and practical deployment for content identification, malware detection, reverse engineering, and optimization. We release the trained Binary BPE tokenizers on HuggingFace, providing a drop-in, open-source foundation for binary-focused language models and context-efficient agentic tools.

The integration of large language models (LLMs) into automated driving systems has opened new possibilities for reasoning and decision-making by transforming complex driving contexts into language-understandable representations. Recent studies demonstrate that fine-tuned LLMs can accurately predict vehicle trajectories and lane-change intentions by gathering and transforming data from surrounding vehicles. However, the robustness of such LLM-based prediction models for safety-critical driving systems remains unexplored, despite the increasing concerns about the trustworthiness of LLMs. This study addresses this gap by conducting a systematic vulnerability analysis of LLM-enabled vehicle trajectory prediction. We propose a one-feature differential evolution attack that perturbs a single kinematic feature of surrounding vehicles within the LLM's input prompts under a black-box setting. Experiments on the highD dataset reveal that even minor, physically plausible perturbations can significantly disrupt model outputs, underscoring the susceptibility of LLM-based predictors to adversarial manipulation. Further analyses reveal a trade-off between accuracy and robustness, examine the failure mechanism, and explore potential mitigation solutions. The findings provide the very first insights into adversarial vulnerabilities of LLM-driven automated vehicle models in the context of vehicular interactions and highlight the need for robustness-oriented design in future LLM-based intelligent transportation systems.

As AI models are trained on ever-expanding datasets, the ability to remove the influence of specific data from trained models has become essential for privacy protection and regulatory compliance. Unlearning addresses this challenge by selectively removing parametric knowledge from the trained models without retraining from scratch, which is critical for resource-intensive models such as Large Language Models (LLMs). Existing unlearning methods often degrade model performance by removing more information than necessary when attempting to ”forget” specific data. We introduce Forgetting-MarI, an LLM unlearning framework that provably removes only the additional (marginal) information contributed by the data to be unlearned, while preserving the information supported by the data to be retained. By penalizing marginal information, our method yields an explicit upper bound on the unlearn dataset's residual influence in the trained models, providing provable undetectability. Extensive experiments confirm that our approach outperforms current state-of-the-art unlearning methods, delivering reliable forgetting and better preserved general model performance across diverse benchmarks. This advancement represents an important step toward making AI systems more controllable and compliant with privacy and copyright regulations without compromising their effectiveness.

Graph machine learning has advanced rapidly in tasks such as link prediction, anomaly detection, and node classification. As models scale up, pretrained graph models have become valuable intellectual assets because they encode extensive computation and domain expertise. Building on these advances, Graph Foundation Models (GFMs) mark a major step forward by jointly pretraining graph and text encoders on massive and diverse data. This unifies structural and semantic understanding, enables zero-shot inference, and supports applications such as fraud detection and biomedical analysis. However, the high pretraining cost and broad cross-domain knowledge in GFMs also make them attractive targets for model extraction attacks (MEAs). Prior work has focused only on small graph neural networks trained on a single graph, leaving the security implications for large-scale and multimodal GFMs largely unexplored. This paper presents the first systematic study of MEAs against GFMs. We formalize a black-box threat model and define six practical attack scenarios covering domain-level and graph-specific extraction goals, architectural mismatch, limited query budgets, partial node access, and training data discrepancies. To instantiate these attacks, we introduce a lightweight extraction method that trains an attacker encoder using supervised regression of graph embeddings. Even without contrastive pretraining data, this method learns an encoder that stays aligned with the victim text encoder and preserves its zero-shot inference ability on unseen graphs. Experiments on seven datasets show that the attacker can approximate the victim model using only a tiny fraction of its original training cost, with almost no loss in accuracy. These findings reveal that GFMs greatly expand the MEA surface and highlight the need for deployment-aware security defenses in large-scale graph learning systems.

Large language models (LLMs) have recently shown strong potential in vulnerability detection (VD). However, accurately detecting vulnerabilities in real-world repositories requires reasoning over complex contextual interactions. Existing LLM-based VD approaches remain limited because current datasets lack complete contextual information and high-quality reasoning supervision, while existing optimization methods primarily rely on coarse outcome-centric supervision signals that fail to model the vulnerability reasoning process. To address these limitations, we first construct ContextVul, a new dataset that augments high-quality function-level vulnerability benchmarks with repository-level contextual information and curated vulnerability reasoning traces. Building upon ContextVul, we introduce a two-stage optimization framework consisting of lightweight cold-start supervised fine-tuning followed by vulnerability-adaptive on-policy optimization (VULPO). VULPO incorporates multidimensional rewards that jointly evaluate vulnerability identification, vulnerability-relevant localization, and causal reasoning quality, along with difficulty-adaptive reward scaling to mitigate reward hacking and improve RL effectiveness. Extensive experiments demonstrate the superiority of VULPO for context-aware VD. Our VULPO-4B, the first specialized vulnerability reasoning LLM, substantially outperforms existing VD baselines, improving Pairwise Pass@1 by 203% relative to Qwen3-4B and achieving competitive performance against a 150% larger-scale LLM, DeepSeek-V3.1.

The rapid proliferation of pretrained models and open repositories has made model merging a convenient yet risky practice, allowing free-riders to combine fine-tuned models into a new multi-capability model without authorization. Such unauthorized model merging not only violates intellectual property rights but also undermines model ownership and accountability. To address this issue, we present MergeGuard, a proactive dual-stage weight protection framework that disrupts merging compatibility while maintaining task fidelity. In the first stage, we redistribute task-relevant information across layers via L2-regularized optimization, ensuring that important gradients are evenly dispersed. In the second stage, we inject structured perturbations to misalign task subspaces, breaking curvature compatibility in the loss landscape. Together, these stages reshape the model's parameter geometry such that merged models collapse into destructive interference while the protected model remains fully functional. Extensive experiments on both vision (ViT-L-14) and language (Llama2, Gemma2, Mistral) models demonstrate that MergeGuard reduces merged model accuracy by up to 90% with less than 1.5% performance loss on the protected model.

Transparency and security are both central to Responsible AI, but they may conflict in adversarial settings. We investigate the strategic effect of transparency for agents through the lens of transferable adversarial example attacks. In transferable adversarial example attacks, attackers maliciously perturb their inputs using surrogate models to fool a defender's target model. These models can be defended or undefended, with both players having to decide which to use. Using a large-scale empirical evaluation of nine attacks across 181 models, we find that attackers are more successful when they match the defender's decision; hence, obscurity could be beneficial to the defender. With game theory, we analyze this trade-off between transparency and security by modeling this problem as both a Nash game and a Stackelberg game, and comparing the expected outcomes. Our analysis confirms that only knowing whether a defender's model is defended or not can sometimes be enough to damage its security. This result serves as an indicator of the general trade-off between transparency and security, suggesting that transparency in AI systems can be at odds with security. Beyond adversarial machine learning, our work illustrates how game-theoretic reasoning can uncover conflicts between transparency and security.