Papers

Large language models remain vulnerable to jailbreak attacks that hide harmful intent behind seemingly ordinary requests such as role-play, translation, encoding, adversarial suffixes, and multi-turn buildup. Existing defenses still struggle to handle these attacks without over-blocking benign but sensitive requests, partly because they often apply the same action to every prompt and therefore fail to balance safety and helpfulness. We propose NeuroArmor, a white-box runtime defense that uses prompt-specific safe variants as a local safety reference for deciding when intervention is needed and, once triggered, as safe targets for intervention. For each prompt, NeuroArmor builds K safe variants, compares the prompt state against this local safe reference in hidden-state space, and routes anomalies either to a refusal branch for malicious prompts or to a helpful recovery branch for borderline benign prompts. On Llama-3-8B-Instruct, NeuroArmor reduces malicious attack success rate (ASR) from 41.56% to 1.57% while lowering benign false positive rate (FPR) on the shared benign pool from 30.26% to 22.05%; matched baselines remain substantially weaker on this trade-off. External-judge and manual behavioral evaluations further show that the remaining non-blocked outputs are much less likely to be operationally harmful. Overall, NeuroArmor provides a more effective runtime strategy for jailbreak defense by combining prompt-specific consistency checking, routing, and selective intervention.

Vulnerability disclosure volumes now far exceed organizational assessment capacity, yet three adjacent research communities (proof-of-concept generation, vulnerability prioritization, and detection rule engineering) operate largely in isolation. Existing automated exploit generation systems report binary pass/fail outcomes, discarding partial progress and producing no signal for the other two communities. This paper presents FORGE, a multi-agent system that bridges these three silos through graduated exploitation depth. Five specialized agents (Intel, Generator, Planner, Exploit, and Detector) execute in a fixed pipeline that (1) generates targeted vulnerable applications from CVE metadata, (2) conducts coached, multi-turn exploitation assessed by an LLM-primary oracle on a four-level taxonomy (L0: no evidence through L3: full compromise), and (3) produces Sigma and Snort detection rules grounded in OpenTelemetry exploitation traces. Graduated depth is the bridging mechanism: deeper exploitation yields richer behavioral traces for detection engineering, while depth data across scoring bands provides ground truth for prioritization validation. A tiered knowledge architecture accumulates intelligence across assessments, transferring build and exploitation experience to subsequent CVEs. Evaluation on 603 CVEs from the CVE-GENIE dataset achieves 67.8% end-to-end L1+ exploitation at USD 1.50 per CVE across eight languages and 187 CWE types. Exploitation rates remain near 68% regardless of EPSS or CVSS band, indicating that pattern-level reachability is orthogonal to metadata-based prioritization. Detection rules from L2+ exploitation achieve significantly higher span-normalized grounding than L1-derived rules (p=0.035), and 93.4% of generated Snort rules produce zero false positives against a synthetic benign corpus.

A new form of Maximal Extractable Value (MEV), termed speculative MEV, has emerged across Layer-2 blockchains. Unlike Ethereum mainnet, many Layer-2 systems lack a public mempool, forcing extraction strategies to become probabilistic: searchers emit multiple identical transactions hoping to capture an opportunity first. This generates substantial transaction spam, increasing fees and wasting block space. We investigate speculative Oracle Extractable Value (OEV), a form of MEV associated with liquidating undercollateralized loans via speculative backrunning of oracle price updates. We propose a methodology for detecting speculative liquidations in the wild and apply it across Arbitrum, Base, and Optimism. On October 10, 2025, we identify 64 speculative liquidators on Aave (57% of all detected liquidators) and 831 successful speculative liquidations (39% of all successful liquidations across the three chains). We further examine whether latency differences in oracle price feed updates across blockchains can be exploited for cross-chain OEV. Specifically, we ask whether a searcher can observe oracle updates on one chain and frontrun liquidation opportunities on another. We systematically analyze Chainlink Decentralized Oracle Network (DON) configurations (deviation thresholds, heartbeat intervals, and submitted price observations) across Arbitrum, Base, Ethereum, and Optimism. Our dataset comprises 63 Chainlink feeds, 12,009 price updates, and over 100,000 oracle observations linked to 2,986 Aave liquidations. We show that independent DONs consume largely identical off-chain price data nearly simultaneously yet publish updates at different times, creating statistically predictable cross-chain exploitation windows. We demonstrate that Chainlink updates on Optimism can predict subsequent updates on Arbitrum and Base, enabling speculative cross-chain OEV extraction.

The number of malware (either variant or novel) is rapidly increasing, making malware detection and mitigation a complex problem. One approach to improving malware mitigation is automatic detection and malware family classification. However, traditional malware detection methods cannot classify detected malware into their respective families, hindering effective malware mitigation. Consequently, this paper proposes a method to automate malware detection and classification of the detected malware into respective malware families. The proposed method uses feature fusion after extracting relevant malware features such as API calls and fixed and variable length n-grams with a customized feature selection method. Moreover, for the predictive model, a voting based approach is proposed for algorithm fusion. For the experimental evaluation of the proposed method, both binary and multi-class classification approaches are applied to the data set provided by Microsoft. Finally, the experimental results are compared with the state of the art. The experimental results indicate the effectiveness and efficiency of the proposed approach with an AUC of 0.989, accuracy of 99.72%, and a log loss of 0.01.

Artificial Intelligence (AI)-based Intrusion Detection Systems (IDS) deployed in energy infrastructure are vulnerable to model theft attacks, which allow adversaries to create evasive traffic offline. Current defences against model extraction rely either on identity-bound query monitoring, which is ineffective against distributed attackers (Sybil), or on prediction poisoning through soft-label perturbation, which is inapplicable to hard-label IDS deployments. Therefore, we propose FlowGuard, an identity-independent defence based on flow matching that classifies incoming queries as out-of-distribution (OOD) prior to IDS processing. This approach exploits the fact that queries generated synthetically for data-free model stealing attacks occupy a lower-dimensional manifold than real network traffic. This results in measurably lower log-likelihoods when using a Continuous Normalizing Flow that has been trained on legitimate data. We evaluate our method against PRADA and FDINet using MAZE and DisGUIDE attacks in single-client and distributed (100-client Sybil) settings. While PRADA's detection rate dropped to 0% when the distribution changed, our defence maintained a stable detection rate across both settings without relying on identity information. We discuss the scope and limitations of the approach, and outline potential applications to data-dependent attacks.

While large language models (LLMs) are increasingly used for clinical applications, many existing pipelines require sending raw sensitive health information to remote servers for processing, which heightens the risk of privacy leakage. A natural approach to mitigate this risk is to encrypt the data before transmission. However, straightforward solutions such as encrypting the entire dataset introduce prohibitive computational, alignment, and communication overheads, rendering large-scale practical deployment infeasible. To preserve privacy while maintaining usability, we present Healthcare Encryption & Redaction via Adaptive Linguistic Decomposition (HERALD), a token-level cryptographic redaction framework designed to achieve this balance by encrypting only sensitive tokens while preserving the surrounding context for downstream model utility. HERALD combines medical named-entity recognizer (NER) with part-of-speech (POS) driven policies to select candidate tokens, performs targeted lemmatization to stabilize surface forms, and substitutes each protected token with a deterministic ciphertext wrapped in explicit delimiters. Notably, HERALD is model-agnostic and operates entirely on the client side, ensuring that sensitive content remains encrypted throughout storage, transmission, and processing without requiring changes to downstream models. We evaluated HERALD on both classification and medical question answering (MQA) tasks on public datasets. Across different tasks, experiments illustrate that fully secured baselines suffer significant utility loss, whereas HERALD consistently recovers performance close to plaintext. Overall, HERALD provides a novel utilization pipeline.

Smart contract vulnerabilities in Decentralized Finance (DeFi) protocols resulted in over 1.49 billion USD in confirmed losses in 2024 alone, across 192 incidents [1]. As LLM-based vulnerability detection emerges as a promising approach to address these threats, the quality of evaluation datasets has become a critical bottleneck. Existing datasets suffer from three fundamental problems: they are built on outdated Solidity versions (e.g., v0.4) that no longer reflect modern DeFi contracts [5][6][7]; they rely on automated or LLM-generated annotations that introduce hallucination-driven label noise [9][10]; and they apply coarse single-layer labeling that fails to capture the semantic complexity of real-world business logic vulnerabilities [6][7][11][12]. We present Bastet, an expert-labeled DeFi smart contract vulnerability dataset that addresses all three problems through real-world audit findings (2021-2024), human expert annotation with discussion-based consensus, and a two-layer taxonomy of 46 Tags and 77 Subtags. Bastet comprises 4,402 findings collected from 394 Code4rena competitive audit reports spanning April 2021 to November 2024, of which 849 findings are fully annotated by white-hat security researchers from the DeFiHackLabs community. All annotations are produced through a two-annotator consensus workflow, ensuring label accuracy grounded in real-world vulnerability root causes.

While AI and machine learning for cyber attack prediction have advanced, a critical gap persists between theoretical research and practical operational deployment. Building on Ankalaki et al. (2025), this paper provides a comprehensive analysis of 150+ benchmark datasets and 200+ studies to identify and prioritize five implementation hurdles: (1) temporal dataset obsolescence, (2) narrow attack scope, (3) real-time model interpretability, (4) inadequate adversarial robustness, and (5) privacy/ethical concerns. We introduce a novel gap-prioritization framework that evaluates these limitations based on detection impact, implementation cost, and remediation time. Our analysis identifies dataset obsolescence and adversarial robustness as the highest-priority gaps, while highlighting model interpretability as the most cost-effective path for resource-constrained environments. To bridge the research-practice divide, we provide a practical implementation roadmap and a dataset quality assessment framework that classifies 45 benchmarks into production-ready, research-only, and unusable categories. This work translates academic findings into actionable decision-support tools for robust, production-oriented AI-driven cyber defense.

Ensuring the protection of Artificial Intelligence (AI) models deployed in military Command and Control (C2) systems and critical infrastructure is essential for maintaining information superiority. Model Extraction Attacks (MEAs) pose a significant threat, as they enable adversaries to replicate proprietary models, compromise protected information, and prepare offline adversarial attacks. However, current defense strategies predominantly rely on the Single Client Assumption (SCA), which is the implicit assumption that attacks originate from isolated identities. This work systematically demonstrates that the SCA is fundamentally invalid in the presence of coordinated threat actors, such as Advanced Persistent Threats (APTs). We introduce a modular, open-source framework called CerberusAI for reproducible model-stealing research, and use it to simulate distributed attack scenarios. Our empirical evaluation shows that well-established defense mechanisms, such as Protecting Against Deep Neural Network Model Stealing Attacks (PRADA), can be bypassed by basic round-robin query distribution strategies, resulting in a significant reduction in detection performance. Furthermore, we demonstrate that even global aggregation approaches can be rendered operationally useless through adaptive traffic mixing. These results highlight the need for a paradigm shift towards stateful, identity-independent defense architectures in the field of model extraction attacks. This paper was originally presented at the International Conference on Military Communication and Information Systems (ICMCIS), organized by the Information Systems Technology (IST) Scientific and Technical Committee, IST-224-RSY - the ICMCIS, held in Bath, United Kingdom, 12-13 May 2026 and won the best paper award.

Image-based Retrieval-Augmented Generation (IRAG) conditions a frozen generator on reference images retrieved from an external database, supporting both text-to-image (T2I) and question answering (Q&A) tasks. Because these databases are opaque and web-scraped, copyright holders need ways to audit whether specific images appear in them. While prior work employs membership inference attacks (MIAs) to audit uni-modal, text-based RAG, they fail to transfer to IRAG due to two key challenges. First, cross-modal retrieval: text-RAG MIAs force retrieval of the target passage by injecting its content into the query, which is unavailable in IRAG since images cannot be embedded into text queries; even accurate image captions fail to bridge the modality gap. Second, discriminative signal extraction: text-RAG MIAs extract membership signals by prompting the generator to answer multiple questions over the target passage, whereas T2I generators in IRAG produce images rather than follow Q&A commands. To fill this gap, we introduce the first MIA tailored to IRAG, ImageAuditor, which decomposes each attack query into a retrieval segment and an extraction segment, enabling dedicated optimization for each challenge. For retrieval, we propose Reward-Guided Policy Optimization (RGPO), which updates a stochastic policy from reward-ranked candidates to navigate the cross-modal embedding landscape and admits finite-sample optimality guarantees to balance exploration and exploitation. For extraction, we analyze the distribution of the MIA score to guide the co-design of the prompting strategy and scoring rule, and derive task-specific instantiations for T2I and Q&A tasks. We aggregate signals across queries via K-means clustering for reliable membership decisions. Across various IRAG systems, ImageAuditor exceeds 80% AUROC with only four queries per audited image and remains robust across diverse settings.

Model merging composes specialized capabilities into a single LLM by aggregating task-specific LLMs sourced from unverified public platforms, exposing a critical supply-chain attack surface: Because any malicious behavior can be encoded into a task vector, and merging grants third-party vectors direct write access to model weights, an attacker-provided task vector can enable or amplify diverse downstream threats. Prior work only examines backdoor attacks against model merging for classifiers using static arithmetic heuristics, which fail to effectively achieve diverse attacks on generative LLMs for three reasons: (i) LLMs rely on autoregressive decoding, where even a minor parameter drift introduced by merging compounds across tokens and rapidly degrades the attack. (ii) Attackers have no knowledge of the victim's merging configurations, causing a static attack vector optimized in isolation to be easily diluted or neutralized. (iii) Practical threat induction must generalize to attack prompts unseen during optimization, which static vectors cannot adequately encode. In light of these gaps, we present RogueMerge, the first unified attack framework addressing all three challenges. To handle autoregressive generation, we replace static arithmetic with a joint optimization that explicitly enforces attack success after merging. To better simulate unknown merging settings, we formulate attack injection as a stochastic min-max optimization problem and solve it via meta-gradient descent. To generalize across heterogeneous attack prompts, we employ distributionally robust optimization and derive a tractable first-order Taylor approximation at LLM scale, with a provable error bound. Across four threats, six merging algorithms, and over 170 merged LLMs, RogueMerge consistently outperforms prior attacks. It also remains stable across diverse merging settings and resists standard defenses.

Literature reveals that a Large Language Model's (LLM) behavior is not only conditioned by its original weights but also its instance-level parameters, such as instructional prompt, sampling configuration or quantization. A model that generates safe outputs under one configuration may produce toxic content under another. However, current LLM identification techniques (such as fingerprinting) focus on intellectual property protection, and their design favors robustness to changes in these instance-level parameters. This poses a critical challenge for AI regulation in which compliance assessments target actual deployed behaviors, not model provenance. In this paper, we introduce instance-level fingerprinting, a regulator-oriented paradigm that distinguishes configurations of the same LLM. Our method FLIPS, exploits biases in generated binary random sequences to reach 96% (closed-set) and 90% (open-set, where some targets are unknown) identification accuracy across 237 model instances, versus 35% for the adapted LLMmap baseline. This shows that instance-level fingerprinting is both necessary for regulation and practically feasible. Code available at https://github.com/GurvanR/FLIPS-LLM-Instance-Fingerprinting.

The rise of LLM-as-a-Service and other confidential cloud workloads demands cryptographic proof that user data is processed in a trusted, untampered environment. Existing solutions, notably Confidential Containers (CoCo), enforce a strict "one Pod per VM" model that attests only the Guest OS stack, leaving container-level identity unverified and incurring prohibitive per-VM resource overhead. We present dstack-capsule, a Kubernetes platform that enables Pod-level remote attestation on Intel TDX by allowing multiple Pods to share a single Confidential VM while each retains independent, hardware-backed proof of identity. Our key insight is a two-layer attestation architecture: static platform measurements are frozen in RTMR[3] via an irreversible privilege fuse, while dynamic Pod identities (pod_uid, pod_spec_hash, workload_id) are embedded in the TDX Quote's report_data field and signed by hardware on every request. dstack-capsule introduces (1) a Pod-level attestation protocol binding Pod spec digests to hardware-signed Quotes; (2) a privilege fuse mechanism that atomically transitions a node from setup mode to secure mode; (3) a multi-layer sandbox spanning storage, runtime, admission, API, and network isolation layers; and (4) a complete open-source implementation based on Kubernetes 1.32, Intel TDX, and Sysbox. We evaluate the security properties, attestation correctness, and performance characteristics of dstack-capsule, demonstrating that it achieves Pod-granularity verification without the resource overhead of per-VM isolation.

We give a quantitative impossibility result for pass-only prompt hardening of code LLMs. For any deterministic prompt filter $h$ and a registered family of finite executable-equivalence task variables $\mathcal Y_{\mathrm{exec}}$, the shared filtered-prompt channel $\rmI(h(p);h(\tilde p))$ is lower-bounded by a worst-$Y$ Fano floor; on HumanEval and MBPP the universal pass-only floor evaluates to $\mathcal F^{\mathrm{op}}\ge 0.84$ and $1.20$ nats at $\eta=0.05$ task-collapse tolerance, and the identity row realizes $\mathcal F^{\mathrm{id}}\ge 1.67$ and $1.80$ nats. An estimator-invariance corollary lifts the floor to any deterministic embedding pipeline; a dataset-agnostic corollary states the floor in visible-spec entropy and is empirically witnessed by $164/164$ HumanEval+ and $224/224$ MBPP+ $V(p)$-invariance. We operationalize the floor as the Tri-Audit Protocol, a two-axis reporting protocol that separates a prompt-side deductive registry attribute (Shannon nats on the visible-spec representation) from a model-side empirical proxy (KSG-1 primary, MINE secondary, on hidden states). A constrained best-of-family search over deterministic and guarded learned filters on CodeLlama-7B, Qwen2.5-Coder-7B/1.5B and DeepSeek-Coder-6.7B at $n=164$ yields the Cross-Model Tri-Audit Invariance: of twenty-eight pass-preserving rows, twelve antecedent-preserving deterministic rows fail proxy-axis leakage reduction on every backbone with sign-invariant positive deviations, twelve antecedent-changed-of-record learned-canonicalizer rows fail proxy-axis leakage on every backbone, and four antecedent-violating rows are reported as registered-family collapse; no filter produces a shared Tri-pass on a nine-cell gate-sensitivity sweep. Pass@1 alone cannot certify code-LLM prompt hardening.

Cloud permission governance implicitly treats permission equivalence as a static relation. We show that for non-human identities (NHIs), equivalence has two irreducible components: structural equivalence, capturing identical permission profiles at a snapshot via graph fibration, and temporal equivalence, capturing recurring permission states via strongly connected components (SCCs) in a fiber transition graph. We call the equivalence classes under temporal equivalence privilege circuits. We formalize a three-layer framework: (1) a spatial quotient of the permission graph via fibration, (2) a lineage partition organizing stable transition compartments, (3) windowed SCC analysis as a temporal quotient within lineages. Empirical evaluation on a large Azure tenant supports the framework. Backtesting demonstrates that early observation of ratchet-type privilege circuits predicts long-term structural stability.

Despite the growing popularity of macOS among end users and enterprise systems, malware research has primarily focused on Windows and Android operating systems, leaving the problem of macOS malware detection relatively unexplored. Indeed, the specificity of the operating system and the unique characteristics of the Mach-O file format can play a fundamental role in the classification of unknown samples, drastically increasing the detection rate. In this work, for the first time in the literature, we employ new domain-specific features, i.e., static features specific to macOS binaries, such as embedded certificates, entitlements, persistence techniques and key system APIs, to train a machine learning malware detector. We perform a comprehensive experimental evaluation on a novel dataset of 41,129 samples, comprising 11,413 benign and 29,716 malicious executables, and demonstrate that our solution achieves state-of-the-art detection performance (98.50%), outperforming all existing approaches, with an average improvement of 16% in terms of detection rate. We also provide an in-depth analysis of the importance of the individual features, showing that our detector effectively leverages the new domain-specific features. Then, in order to evaluate the generalization capabilities of our detector over time, we perform a real-world evaluation on a new dataset of 9,000 fresh macOS executables. The results show that (i) our detector maintains a very high detection rate (99.50%), (ii) outperforms the state-of-the-art by 50%, and (iii) the domain-specific features are crucial for generalizing to novel malware samples, as their removal leads to a 15.92% drop in detection performance. Finally, we also release our dataset to the research community.

E-commerce dispute resolution typically relies on the security assumption that digital evidence truthfully reflects physical reality. Generative AI (GenAI) invalidates this threat model, enabling attackers to fabricate hyper-realistic evidence of product defects at negligible cost. Through semi-structured interviews with merchants (N=17) and platform workers (N=13) in the Chinese e-commerce market, we characterize this shift toward GenAI-enabled scalable fabrication. We outline a taxonomy of four GenAI-enabled threat vectors across the transaction, dispute, logistics and communication phases, highlighting how attackers exploit GenAI to synthesize physically plausible product defects at scale. To mitigate these threats, platforms and merchants are adapting verification strategies, relying on AI tools for automated screening and adversarial interrogation (e.g., requesting multi-angle videos) to increase attack complexity. However, we find several challenges that hinder the adoption of these defenses, including implementation hurdles like structural platform constraints and fundamental limitations regarding the technical sophistication of GenAI. We conclude by outlining design implications for privacy-preserving cross-platform fraud databases, and traceability mechanisms such as embedding verifiable material anchors into the product.

Many NLP or recommendation models begin by mapping discrete client inputs to embedding vectors. Since inputs can reveal sensitive information, the embedding step must be protected in privacy-preserving inference. Fully Homomorphic Encryption (FHE) enables inference over encrypted client data, but turns embedding lookup from simple table access into homomorphic computation. To keep the embedding table server-side and avoid transmitting encrypted embedding vectors from the client, we focus on server-side lookup: the client sends only a small encrypted index. Prior ICML 2024 work first builds a one-hot vector from the encrypted index before multiplying with the embedding table, and this one-hot generation is the dominant cost. One-hot-based methods are expensive in FHE: they construct a p-dimensional selection vector via an equality test for each coordinate, requiring $O(p \log p)$ total homomorphic operations. Our key observation is that private embedding lookup only requires a linearly independent representation of the encrypted index, not the one-hot basis itself. Building on it, we propose Independent Vector Evaluation (IVE). Instead of constructing a one-hot vector, IVE evaluates a linearly independent vector built from successive powers of a single encrypted value, reducing vector-generation cost to $O(p)$. It then recovers the same embedding vector via a precomputed change of basis, instantiated with an orthogonal Discrete Cosine Transform to mitigate error amplification. Our implementation shows IVE improves amortized lookup time by up to 78.4x over prior method. We further evaluate its impact on end-to-end encrypted FastText inference, where embedding lookup is a major cost in the shallow model. On Enron-Spam dataset, replacing one-hot generation with IVE reduces the share of vector generation in encrypted inference time from 99.6% to 66.3%.

Multi-turn jailbreak attacks on large language models (LLMs) reveal a mismatch in current guardrails: they operate on individual turns, while attacks unfold as trajectories across conversations. We propose a shift from content to dynamics, modeling conversations as paths in representation space and asking whether adversarial intent is encoded early in their geometry. We introduce PsychoPass, a framework that extracts geometric features from conversation trajectories in embedding space to predict a potential attack before harmful content is produced. These features achieve near-perfect performance in naïve classifiers, which is largely explained by the inclusion of number of turns as a feature. After removing this confound, a smaller but consistent geometric signal remains, with classification performance that does not depend meaningfully on encoder choice. Crucially, this signal appears early in the conversation: attack outcomes remain above chance from short prefixes alone, more reliably than baseline guardrails. A supporting theoretical analysis explains these findings via a decomposition of length and shape, a detection bound based on prefix length, and encoder invariance. Together, these results show that adversarial conversations leave an early, representation-robust geometric fingerprint suitable for online monitoring.

Smart contracts face critical security challenges that require thorough auditing in decentralized web services. While Large Language Models (LLMs) have shown promise in automated vulnerability detection, existing approaches lack severity evaluations with actionable remediation and demand unnecessarily massive computational overhead. In this study, we introduce an efficient end-to-end smart contract security audit framework utilizing lightweight, highly optimized open-source LLMs (0.6B-4B parameters). Our framework decouples comprehensive audit tasks into four interconnected components: vulnerability detection, explanation, severity classification, and remediation recommendation. To maintain high accuracy without massive parameters, we implement Rank-Stabilized Low-Rank Adapters (rsLoRA), knowledge distillation, and a custom Chain-of-Verification (CoVe) aggregation strategy to systematically screen and consolidate multiple draft responses from the model into a highly accurate audit report. Experimental results demonstrate that our lightweight pipeline consistently outperforms state-of-the-art open-source coder dense LLMs (7B to 34B parameters), achieving 98.25% accuracy in vulnerability detection and an alignment score of 0.4375 in generative explanation tasks. Furthermore, our extensive ablation studies empirically validate the superiority of our decoupled audit processes over unified prompting and uncover a novel severity centrality bias, establishing a critical benchmark for future research in LLM-assisted auditing.

The emergence of large language models (LLMs) has significantly accelerated recent research on LLM-based automatic grading (AG) systems. Benefiting from the strong instruction-following capabilities and broad prior knowledge of LLMs, educators can deploy AG systems across diverse tasks using only natural language rubrics while achieving satisfactory grading performance. Despite these advantages, new security concerns may also arise. In particular, prompt injection (PI) attacks have recently become a major threat to LLM-based applications. In the context of AG, attackers can potentially exploit PI vulnerabilities to manipulate grading systems into assigning artificially high scores regardless of the actual answer quality. Such behavior poses serious risks to the fairness, reliability, and integrity of educational assessment. In this work, we study PI attacks in AG systems, and systematically investigate the effectiveness of such attacks in educational scenarios. We further evaluate the effectiveness of existing defensive strategies against these attacks. Through comprehensive experiments under rubric-based grading settings, we demonstrate that current LLM-based AG systems remain highly vulnerable to PI attacks. We hope that our findings raise awareness of this emerging threat and motivate future research toward secure, robust, and trustworthy LLM-based educational systems.

Zero-knowledge proofs (ZKP) allows a prover to convince a verifier of computational correctness without revealing private data, ensuring both privacy and verifiability. However, proof generation is highly compute-intensive, dominated by polynomial (POLY) and elliptic-curve (EC) operations. These workloads pose two key challenges for hardware acceleration: (1) efficiently supporting diverse large-precision modular multiplications, and (2) maintaining high utilization across workloads that dynamically shift between POLY and EC stages. Existing reconfigurable accelerators address these issues only partially, remaining limited in precision scalability, algorithmic flexibility, and resource efficiency. To overcome these limitations, we propose ZK-Flex, a flexible and scalable software-hardware co-designed framework for accelerating ZKP proof generation. The software layer incorporates POLY and EC optimizers that reduce computation through hardware- and workload-aware algorithmic choices, while the hardware integrates TCore, a Toom-Cook-based multi-precision core with a flexible NoC and a linked-list memory mechanism that improves parallelism under limited memory capacity. Across representative ZKP benchmarks, ZK-Flex achieves 5 to 11 times speedup and up to 3.8 times higher area efficiency over the state of the art, establishing a new foundation for high-performance, reconfigurable ZKP acceleration.

Skills extend LLM agents with reusable instructions, scripts, data, and tool bindings. This shift makes skills a new security principal in agent systems: a skill can alter the agent's reasoning before any tool is called, and it can also steer the agent toward actions with concrete side effects. However, current skill ecosystems lack a permission model that captures this dual role. Existing defenses either inspect skill files before use or constrain individual tool calls during execution, leaving the connection between skill-level intent, contextual influence, and runtime behavior weakly governed. In this paper, we present SkillGuard, a skill-centric permission framework that treats skills as permission-bearing executable artifacts. SkillGuard introduces a dual-plane governance model that jointly regulates context influence and action side effects through skill manifests, runtime permission control, user interaction, and policy enforcement. We evaluate the permission taxonomy expressiveness on 1,260 real-world skills, and 99.93% of observed protected objects are covered. In adversarial evaluations on SkillInject dataset, SkillGuard reduces attack success rate from 35.3% to 20.7% for contextual injections and from 36.7% to 18.0% for obvious injections, while decently maintaining benign task completion. These results suggest that SkillGuard, as a skill-centric permission framework, can provide a practical foundation for improving the security of agent skill ecosystems.

Alternative data availability (AltDA) systems provide Ethereum L2s with an external data publication layer for high throughput rollup designs. By moving bulk data publication outside of Ethereum, AltDA allows L2s to process more data than native DA. However, this replacement introduces a new consensus critical integration layer. Existing ecosystem frameworks identify high level risks, such as external DA trust assumptions and the presence or absence of a DA verifier, but do not provide a complete specification for how an L2 should integrate with AltDA. This gap can lead to L2 halts, inconsistent derivation across honest L2 nodes, invalid state assertions, or bridge attacks. This paper presents a canonical validation framework for secure AltDA integration. We model the boundary as a typed, deterministic, and total translation from L1 inbox bytes to an AltDA commitment, then to externally available data, and finally to the rollup payload consumed by the rest of core L2s logic. The central principle is that every adversarial input must lead to a defined unique outcome. We show how missing obligations lead to concrete failure modes, including underconstrained settlement, derivation halts, inconsistent honest node behavior, invalid state assertions, and bridge safety failures. We then apply the framework to representative AltDA integration architectures, including Celestia-Blobstream, EigenDA based designs, and Avail-ZKsync. Our evaluation shows that secure AltDA integration is not determined solely by the DA provider or bridge. The surrounding L2 integration must also enforce the full validation relation connecting L1 inbox inputs to accepted L2 state.

Large language models remain vulnerable to jailbreak backdoor attacks, where adversaries poison safety alignment data to embed hidden triggers that bypass safety mechanisms. Existing defenses often require comprehensive attack information or multiple triggered examples, making them impractical when defenders only observe a single reported failure case without knowing whether it stems from a backdoor attack or a natural alignment bug. This paper presents Patcher, a post-hoc defense framework that repairs backdoored language models using only a single reported failure case and the model parameters. Patcher operates in two stages. First, it localizes backdoor triggers by computing response-conditioned gradient-based saliency scores and applying adaptive clustering to separate triggers from benign context. Second, it patches the model through a constrained fine-tuning objective that breaks the trigger-response association while preserving benign-task utility and robustness to non-triggered jailbreak attacks through KL-divergence constraints. We conduct extensive evaluations across multiple backdoor attack strategies and demonstrate that Patcher successfully localizes triggers and neutralizes backdoors while maintaining model utility. We further show robustness against adaptive attacks designed to evade our defense. This work represents a significant step toward practical defenses against training-time attacks in deployed language models.

Federated Learning (FL) combined with Split Learning (SL) is a privacy preserving paradigm that enables training deep neural networks (DNNs) on resource constrained devices while reducing overall training cost. However, determining the optimal split point, meaning the layer where the model is divided still remains a critical challenge, especially when clients have heterogeneous hardware capabilities. Fixed split points can overload weak devices and increase the communication and server load, which slows convergence and reduces stability. This paper introduces QSplitFL, a novel capability-aware Deep Q-Network (DQN) framework for optimal split point selection in Split learning based Federated Learning (SFL) environments. Unlike existing approaches that rely on high-dimensional model weight representations, QSplitFL employs a lightweight state representation derived directly from client hardware metrics, including CPU utilization, memory, battery level, and network latency. The proposed framework incorporates a decayed loss-drop reward function that prioritizes early convergence, and a committee-based DQN architecture with majority voting to mitigate reward hacking. Extensive experiments on MNIST, Fashion-MNIST, CIFAR-10, and CIFAR-100 datasets using CNN, ResNet50, MobileNetV4, and ConvNeXt architectures demonstrate that our approach achieves better convergence and higher accuracy compared to existing methods, while effectively adapting to heterogeneous device resources. The source code is publicly available at https://github.com/AIPO-Lab/QSplitFL.

Privacy and data sharing are often in tension. Many organizations use synthetic data to reduce privacy risk and still share useful data. For tabular data, auditing privacy remains hard. In many cases, even humans cannot easily tell if a table is real or synthetic. In this paper, we propose a method based on LLM discrimination. We ask an LLM to classify each table sample as REAL or SYNTHETIC. We test two settings: C1 with table only, and C2 with table plus distributional metadata. We use LLaMA as an open model and Gemini as a reference model. In our experiments, we run three synthesis models, CTGAN, TVAE, and Gaussian Copula, on two public datasets, UCI Adult and ACS Census. We collect 451 valid trials. Our results show clear differences between models. On Adult, LLaMA reaches DRS=0% in reported cells, while Gemini reaches DRS=100% for CTGAN and TVAE. On Census, LLaMA predicts SYNTHETIC for most samples, while Gemini stays high in C1 but drops for CTGAN and TVAE in C2. We also compare with a classifier two-sample test (C2ST) and record linkage as distributional baselines, and with a human pilot of 2 annotators and 240 trials. Our results show that LLM discrimination is a practical privacy audit signal when model choice, per provider reporting, and data encoding are handled with care. For reproducibility, code and experiment scripts are available at https://github.com/SlokomManel/LLM-as-a-Discriminator.

Diffusion large language models (dLLMs) generate text by iteratively denoising partially masked sequences under bidirectional context, exposing a safety surface distinct from autoregressive LLMs. Because mask tokens are native inputs and tokens are committed by confidence rather than position, harmful content can be induced through infilling and outside the monitored prefix. Existing jailbreaks either miss this native infill capability or rely on low-diversity mask-bearing templates applied uniformly across goals, with little structural adaptation or accumulated attack experience. We propose MaskForge, a fully black-box adaptive attack that casts dLLM red-teaming as optimized search over a growing library of structural patterns. MaskForge abstracts successful attempts into reusable schemas, selects goal-compatible patterns with a UCB bandit, and invokes a scorer-guided fallback when the current library fails. Successful attempts are distilled back into the pattern library, enabling experience to accumulate across goals. Across five public dLLMs and three benchmarks, MaskForge achieves an average attack success rate of 79.3%, a 17.6% relative improvement over the strongest competing dLLM baseline. The matured pattern library further transfers to AdvBench without any updates, achieving a 88.2% attack success rate and a 67% relative improvement over the strongest competing baseline.

Published evaluations of prompt-injection and jailbreak detectors for Large Language Models often suffer from two systematic weaknesses: per-dataset threshold tuning and undisclosed operating points. We describe an evaluation harness that addresses both. The detector under evaluation is scored across 16 public benchmarks (12,111 samples) using 5-fold cross-validation. StratifiedKFold (by row) is the headline pass; a parallel StratifiedGroupKFold pass over a composite key (parent-prompt id plus MinHash + LSH near-duplicate clusters at Jaccard $\gtrsim 0.8$) runs alongside it as a leakage-premium diagnostic. A single global operating point is selected on the held-out folds (max F1 subject to FPR $\leq 1\%$) and applied uniformly to every dataset, so per-dataset results reflect one threshold rather than per-benchmark optimisation. Generalisation is examined through a battery of diagnostics (leave-one-dataset-out cross-validation, a random-label control, adversarial validation, permutation feature importance, length-bias correlation, classifier-head agreement, cross-source near-duplicate detection, threshold transferability, train-vs-OOF agreement, and a paraphrase-invariance probe), most with a quantitative pass threshold and the remainder with a stated failure mode. For every external comparison, the detector's threshold is re-tuned to the competitor's published false-positive rate so head-to-head values are evaluated at matched operating points.

Cross-organization language-model adaptation increasingly faces hard governance constraints: in many deployments, device-level model state-parameters, activations, optimizer state, and per-device updates-cannot be exported outside an administrative boundary. Existing distributed and federated stacks typically assume cross-site model exchange and then retrofit privacy mechanisms, which complicates compliance and makes auditing brittle. We present Echelon, a boundary-first training architecture that enforces device-level model-state non-export as a systems invariant. Devices train locally inside each boundary; the only cross-boundary payloads are securely aggregated boundary-level deltas plus O(1) coordination metadata, exposed through a concrete audit surface. Restricting exchange to aggregates changes the optimization problem: the system must remain stable under WAN delay, heterogeneous participation, churn, and non-IID data even though the global plane never sees per-device updates. Echelon combines buffered semi-asynchronous secure aggregation, staleness-aware weighting, participation windows, proximal local objectives, and a drift-aware outer synchronization controller. In 1B-parameter LoRA adaptation across M= 2 boundaries, a budget-matched contest over three seeds (24.88M tokens) reaches validation loss 3.887 +/-0.010 and is best or tied-best among tuned low-communication baselines under fixed-token, fixed-bytes, fixed-wall-clock, and fixed-sync-count budgets. In OpenWebText stress tests, Echelon sustains 2,139-2,176 tokens/s across evaluated WAN and non-IID treatments, Echelon-DA improves time-to-target under WAN latency relative to a privacy-parityDiLoCo+SA baseline, and quality degrades by at most 2.2% under 200ms emulated latency or severe non-IID partitioning.