Papers

Decentralized Autonomous Organizations (DAOs) are inclined explore Small Language Models (SLMs) as edge-native constitutional firewalls to vet proposals and mitigate semantic social engineering. While scaling inference-time compute (System 2) enhances formal logic, its efficacy in highly adversarial, cryptoeconomic governance environments remains underexplored. To address this, we introduce Sentinel-Bench, an 840-inference empirical framework executing a strict intra-model ablation on Qwen-3.5-9B. By toggling latent reasoning across frozen weights, we isolate the impact of inference-time compute against an adversarial Optimism DAO dataset. Our findings reveal a severe compute-accuracy inversion. The autoregressive baseline (System 1) achieved 100% adversarial robustness, 100% juridical consistency, and state finality in under 13 seconds. Conversely, System 2 reasoning introduced catastrophic instability, fundamentally driven by a 26.7% Reasoning Non-Convergence (cognitive collapse) rate. This collapse degraded trial-to-trial consensus stability to 72.6% and imposed a 17x latency overhead, introducing critical vulnerabilities to Governance Extractable Value (GEV) and hardware centralization. While rare (1.5% of adversarial trials), we empirically captured "Reasoning-Induced Sycophancy," where the model generated significantly longer internal monologues (averaging 25,750 characters) to rationalize failing the adversarial trap. We conclude that for edge-native SLMs operating under Byzantine Fault Tolerance (BFT) constraints, System 1 parameterized intuition is structurally and economically superior to System 2 iterative deliberation for decentralized consensus. Code and Dataset: https://github.com/smarizvi110/sentinel-bench

AI agents increasingly call external tools (file system, network, APIs) through the Model Context Protocol (MCP). These tool calls are the agent's syscalls: privileged operations with side effects on shared state, yet today's safety enforcement lives entirely in userspace, where a 10-line script can bypass it. I propose Governed MCP, a kernel-resident tool governance gateway built on a logit-based safety primitive (ProbeLogits). The gateway interposes on every MCP tool call in a 6-layer pipeline: schema validation, trust tier, rate limit, adversarial pre-filter, a ProbeLogits semantic gate (the load-bearing check), and constitutional policy match, with a Blake3-hashed audit chain. I implement Governed MCP in Anima OS, a bare-metal x86-64 kernel in ~286,000 lines of Rust. The five non-inference layers plus the audit append cost a measured 11.3 us per call; the ProbeLogits gate (one probe-prompt prefill plus a single logit read) costs 332-556 ms per classification across Qwen2.5-7B, Llama-3-8B, and Mistral-7B, 2.4-3.4x faster than a Llama Guard 3 pass on the same hardware. A silicon-measured ablation shows that removing the ProbeLogits layer collapses F1 from 0.789 to 0.357 (delta-F1 = -0.432): hand-rule firewalling alone is insufficient. Every WASM-to-system host function and every registered MCP tool is mediated by the kernel gateway, so the 10-line userspace bypass that defeats existing guardrail libraries is structurally impossible; a disclosed set of ring-3 syscall paths remains ungated pending future work. Multi-model validation across three architectures (HarmBench 98-99% non-copyright block, XSTest 98.5-100% unsafe recall, ToxicChat parity with Llama Guard 3) shows the underlying primitive is architecture-agnostic. Governed MCP demonstrates that tool-call governance is feasible as an OS primitive, not just an application-layer concern.

We present enclawed, a hard-fork hardening framework built on the OpenClaw AI assistant gateway. enclawed targets deployments that need attestable peer trust, deny-by-default external connectivity, signed-module loading, and a tamper-evident audit trail – typically regulated industries (financial services, healthcare, defense, government). The framework ships in two flavors: an open flavor preserving OpenClaw compatibility while emitting audit, classification, and data-loss-prevention (DLP) signals, and an enclaved flavor activating strict allowlists, FIPS cryptographic-module assertion, mandatory manifest signature verification, and high-assurance peer attestation for the Model Context Protocol. The classification ladder is data-driven: deployers pick from five built-in presets or supply their own JSON. We ship a 356-case test suite (261 unit + 95 adversarial pen-tests) covering tamper detection, signature forgery, egress bypass, audit-log truncation, trust-root mutation, DLP evasion, prompt injection, code injection, and biconditional admission for net-capable extensions; real-time human-in-the-loop control; a memory-bounded transaction buffer with rollback; strict-mode TypeScript typecheck; and a CI workflow. The biconditional extension-admission gate extends the skill trust schema to non-skill extensions. The four-level verification lattice is now closed at the top: four skill-formal-* primitives plus a CLI produce a signed proof-carrying bundle the runtime re-checks at load, raising a skill from tested to formal via static effect-containment, refinement-typed dispatch, and bounded model checking. enclawed is a hardening framework, not an accredited certification; hardware, validated crypto, facilities, and assessor sign-off remain the deployer's responsibility.

Privacy-preserving machine learning (PPML) has become increasingly important in applications where sensitive data must remain confidential. Homomorphic Encryption (HE) enables computation directly on encrypted data, allowing neural network inference without revealing raw inputs. While prior works have largely focused on inference over a single encrypted image, batch processing of encrypted inputs lags behind, despite being critical for high-throughput inference scenarios and training-oriented workloads. In this work, we address this gap by developing optimized algorithms for batched HE-friendly neural networks. We also introduced a pipeline architecture designed to maximize resource efficiency for different batch size execution. We implemented these algorithms and evaluated our work using HE-friendly ResNet-20 and ResNet-34 models on encrypted CIFAR-10 and CIFAR-100 datasets, respectively. For ResNet-20, our approach achieves an amortized inference time of 8.86 seconds per image when processing a batch of 512 encrypted images, with a peak memory usage of 98.96 GB. These results represent a 1.78x runtime improvement and a 3.74x reduction in memory usage compared to the state-of-the-art design. For the deeper ResNet-34 model, we achieve an amortized inference time of 28.14 on a batch of 256 encrypted images using 246.78GB of RAM

Timing side-channel attacks exploit variations in program execution time to recover sensitive information. Cryptographic implementations are especially vulnerable to these attacks, since even small timing differences in operations such as modular exponentiation or key comparisons can be exploited to extract highly sensitive information, such as secret keys. To mitigate this threat, implementations of programs that handle sensitive information are often expected to adhere to constant-time principles, ensuring that execution behavior does not depend on secret inputs. However, validating the constant-time property of programs remains a major challenge in cryptography development. Formal method approaches to verify constant-time implementations rely on abstractions that often fail to capture real execution behavior, while timing-based measurement techniques are highly sensitive to noise from other programs and even hardware environments. In this work, we propose a novel approach for verifying constant-time programs based on dynamic analysis of low-level execution traces. Our method measures instruction sequences across multiple input values for any given binary and targeted function. Any variations in the instruction mix distribution for any given pair of traces indicate a deviation from the constant-time principle and behavior. We developed an open-source tool called DALC-CT, for the constant-time verification of programs using this approach. We evaluated it on a set of well-known constant-time and non-constant-time examples, achieving a perfect detection of issues. Our results demonstrate that analyzing the logical execution of programs via instruction trace comparisons provides a lightweight and reliable way to verify the constant-time property of programs.

Academic examination systems worldwide continue to rely on centralised, opaque record-keeping that is often vulnerable to credential forgery, result tampering, examiner bias, and the absence of transparent re-evaluation pathways. Existing blockchain-based approaches in education focus predominantly on post-hoc certificate storage or online-only examination portals, leaving the complete onsite examination lifecycle, from conducting exams through scrutiny, largely unaddressed. This paper proposes ParikkhaChain, a blockchain-based framework that covers the entire examination lifecycle of an onsite examination system with three distinguishing contributions: (i) anonymous script evaluation through cryptographic hashing of answer scripts before examiner access, thereby eliminating identity-based bias; (ii) a transparent evaluation and scrutiny workflow backed by an immutable on-chain audit trail that records every mark submission and grade revision; and (iii) inclusion of privacy-preserving verification using zero-knowledge proofs and off-chain storage mechanisms. The system is architected around four Solidity smart contracts deployed on the Ethereum blockchain. The proposed architecture is the first initiative to our knowledge to support physical examination process, anonymous marking, and re-evaluation transparency. We successfully simulate full exam cycles of an onsite exam to grade-sheet generation using a working prototype on a large scale of 100 courses and hundreds of teachers and students. The experimental results show that the system can manage online examinations of hundreds of courses, students and faculties efficiently with great throughput, low storage, and transaction cost. Our codebase is available in open source form at https://github.com/AhmedRumi/CSE6608-ParikkhaChain

Multi-turn jailbreak attacks progressively erode LLM safety alignment across seemingly innocuous conversation turns, achieving success rates exceeding 90% against state-of-the-art models. Existing alignment-based and guardrail methods suffer from three key limitations: they require costly weight modification, evaluate each turn independently without modeling cumulative safety erosion, and detect attacks only after harmful content has been generated. To address these limitations, we first formulate the proactive early jailbreak detection problem with a new metric, detection lead, that measures how early an attack can be detected before the LLM complies. We then propose SAFEDREAM, a lightweight world-model-based framework that operates as an external module without modifying the LLM's weights. SAFEDREAM introduces three components: (1) a safety state world model that encodes LLM hidden states into a compact safety representation and predicts how it evolves across turns, (2) CUSUM detection that accumulates weak per-turn risk signals into reliable evidence, and (3) contrastive imagination that simultaneously rolls out attack and benign futures in latent space to issue early alarms before jailbreaks occur. On three multi-turn jailbreak benchmarks (XGuard-Train, SafeDialBench, SafeMTData) against 8 baselines, SAFEDREAM achieves the best detection timeliness across all benchmarks (1.06-1.20 turns before compliance) while maintaining competitive false positive rates and outperforming baselines in detection quality.

Modern AI agents routinely depend on secrets such as API keys and SSH credentials, yet the dominant deployment model still exposes those secrets directly to the agent process through environment variables, local files, or forwarding sockets. This design fails against prompt injection, tool misuse, and model-controlled exfiltration because the agent can both use and reveal the same bearer credential. We present CapSeal, a capability-sealed secret mediation architecture that replaces direct secret access with constrained invocations through a local trusted broker. CapSeal combines capability issuance, schema-constrained HTTP execution, broker-executed SSH actions, anti-replay session binding, policy evaluation, and tamper-evident audit trails. We describe a Rust prototype integrated with an MCP-facing adapter, formulate conditional security goals for non-disclosure, constrained use, replay resistance, and auditability, and define an evaluation plan spanning prompt injection, tool misuse, and SSH abuse. The resulting system reframes secret handling for agentic systems from handing the model a key to granting the model a narrowly scoped, non-exportable action capability.

Ransomware detection systems increasingly rely on behavior-based machine learning to address evolving attack strategies. However, emerging privacy compliance, data governance, and responsible AI deployment demand not only accurate detection but also the ability to efficiently remove the influence of specific training samples without retraining the models from scratch. In this study, we present a privacy-aware machine unlearning evaluation framework for reinforcement learning (RL)-based ransomware detection built on Sharded, Isolated, Sliced, and Aggregated (SISA) training. The framework enables efficient data deletion by retraining only the affected model shards rather than the entire detector, reducing the retraining cost while preserving detection performance. We conduct a controlled comparative study using value-based RL agents, including Deep Q-Network (DQN) and Double Deep Q-Network (DDQN), under identical experimental settings with a cost-sensitive reward design and 5-fold cross-validation on Windows 11 ransomware dataset. Detection confidence is evaluated using a continuous Q-score margin, enabling ROC-AUC analysis beyond binary predictions. For unlearning, the dataset is partitioned into five shards with majority-vote aggregation, and a fast-unlearning path is evaluated by deleting 5% of the samples from a single shard and retraining only that shard. Results show that SISA-based unlearning incurs negligible utility degradation (<= 0.05 percent F1 drop) while substantially reducing retraining time relative to full SISA retraining. DDQN exhibits slightly improved stability and lower utility loss than DQN, while both agents maintain near identical in-distribution performance after unlearning. These findings indicate that SISA provides an efficient unlearning mechanism for RL-based ransomware detection, supporting privacy-aware deployment without compromising security effectiveness.

As Cyber-Physical Systems (CPS) become increasingly pervasive and autonomous, ensuring the resilience of their embedded logic is critical to maintaining safety and integrity. Among the most stealthy and damaging threats are non-invasive fault injection attacks, where hardware-level disturbances propagate into software execution and compromise control logic. In this paper, we investigate the susceptibility of Unmanned Aerial Vehicle (UAV) autopilot fail-safe mechanisms to voltage glitch fault injection. We introduce a dual evaluation approach: software-based fault simulation using ARMORY and hardware-based experiments with a voltage glitching platform (Chip-Whisperer), applying controlled and timely faults to an STM32 microcontroller running UAV-Autopilot fail-safe logic. Our targeted analysis of specific fail-safe modes uncovers timing-sensitive vulnerabilities that can suppress or alter safety responses, such as disabling emergency failsafe activation at critical moments, potentially enabling UAV hijacking. Furthermore, we validate software-based fault injection results against real hardware behavior, demonstrating how simulated attacks translate into tangible risks for CPS security and reliability.

Large language models write production code, and yet they routinely introduce well-known vulnerabilities. We show that this is not a knowledge deficit: the same models that generate insecure code, correctly identify and explain the vulnerability when asked directly, this is a gap we call the Format-Reliability Gap. Mechanistic analysis reveals the cause: security representations are encoded from the earliest layers but remain computationally inert until the final layer, where format-compliance demands compete with them. Because the failure is localized to a single layer, per-vulnerability steering vectors reduce insecure generation by up to 74% with negligible overhead. The mechanism and the fix generalize across five models, three architecture families, and six vulnerability types, suggesting insecure code generation is an interpretability problem, not a training artifact.

The modern cryptographic primitives are known to generate large volumes of sequential data like keystreams, ciphertext blocks, and hash outputs. Traditional cryptgraphic evaluation methods rely primarily on statistical randomness tests and algebraic cryptanalysis techniques. This paper introduces the concept of Stringology-Based Cryptology (SBC), which applies classical string processing and pattern matching techniques to analyze structural properties of cryptographic outputs. By interpreting cryptographic outputs as symbolic sequences, stringology algorithms can be used to detect pattern recurrence, substring distributions, and structural correlations. In addition, the paper demonstrate how pattern frequency analysis and substring recurrence metrics can be applied to evaluate keystream outputs generated by stream ciphers. Experimental results illustrate that SBC analysis provides complementary insights into structural characteristics of cryptographic sequences and may support future research in structural cryptanalysis and cryptographic evaluation

Fine-tuning on benign data is known to degrade safety alignment in text and vision LLMs, but whether distinct input properties drive this vulnerability differently remains unclear. Audio introduces a richer problem where benign samples can neighbor harmful content through what is said or how it sounds. We present the first systematic study of benign fine-tuning safety in Audio LLMs, evaluating three state-of-the-art models with a proximity-based framework that decomposes embedding-space distance into semantic, acoustic, and mixed axes. We find that the dominant vulnerability axis is architecture-conditioned, determined by how each model's encoder and projector transform audio into the backbone LLM's input space. Across three models, benign fine-tuning elevates Jailbreak Success Rate (JSR) from single digits to as high as 87%, with the most damaging axis shifting from semantic to acoustic proximity depending on encoder design. Mechanistically, fine-tuning selectively suppresses late-layer refusal circuits while frozen encoders preserve upstream representations: the model still detects harmful content but stops refusing, a recognition-refusal dissociation. Two practical defenses, filtering training data to maximize distance from harmful embeddings and a textual system prompt at inference, reduce JSR to near-zero without architectural modification. These findings show that safety evaluation should account for modality and architecture, while highlighting Audio LLMs as a useful testbed for understanding alignment fragility.

Large language models (LLMs) are increasingly deployed in high-stakes domains, yet a unified treatment of their overlapping safety challenges remains lacking. We present SafeLM, a framework that jointly addresses four pillars of LLM safety: privacy, security, misinformation, and adversarial robustness. SafeLM combines federated training with gradient smartification and Paillier encryption for privacy, integrates defenses against training and inference-time attacks, employs contrastive grounding with calibrated decoding to reduce hallucinations, and introduces alignment-aware binarized aggregation to enhance robustness while maintaining bounded reconstruction quality. Across benchmarks on factuality, toxicity, and membership inference, SafeLM achieves 98.0% harmful content detection accuracy, reduces communication by 96.9%, and lowers gradient inversion PSNR from 31.7 dB to 15.1 dB. Ablations show that each component contributes independently, whereas their integration yields a strong privacy utility efficiency trade-off for deploying trustworthy LLMs.

Light clients are essential for scalable blockchain systems because they verify data availability without downloading full blocks. In data availability sampling based systems, sampled cells are retrieved from a peer-to-peer network and verified against cryptographic commitments. A common deployment pattern associates each sampled cell with an independent Kate-Zaverucha-Goldberg (KZG) proof, creating substantial cumulative bandwidth, storage, and verification overhead. This paper studies polynomial multiproofs (PMP) as a mechanism for reducing these costs in blockchain light clients. We present a design in which multiple sampled cell evaluations are verified using a single aggregated proof over a shared evaluation micro-domain and describe the corresponding changes to proof generation, dissemination, retrieval, and verification in a peer-to-peer light-client stack. We instantiate and evaluate the design in Avail, a modular data availability layer for blockchains, as a case study. The results show lower proof bytes, lower verifier CPU and memory usage, and deployment-level infrastructure cost reductions of up to 45% relative to a per-cell baseline, while also clarifying the trade-offs introduced by grouped retrieval.

The emergence of writable, cross-session persistent memory in LLM agents introduces a qualitatively different threat landscape from conventional input-centric security concerns, characterized by three properties: persistence, statefulness, and propagation. To systematically characterize this landscape, we propose a Memory Lifecycle Framework that organizes attacks, defenses, and their cross-phase dependencies along two axes: six lifecycle phases (Write, Store, Retrieve, Execute, Share & Propagate, Forget & Rollback) and four security objectives (Integrity, Confidentiality, Availability, Governance). This analysis in turn exposes the need for formal security guarantees at the system level, motivating Verifiable Memory Governance (VMG), a framework of five architectural primitives that specifies what verifiable mechanisms a long-term-memory system must provide to maintain auditable, recoverable control over its memory state. Our analysis indicates that robust Long-Term Memory (LTM) security cannot be retrofitted at retrieval or execution time alone, but must be anchored in storage-time provenance, versioning, and policy-aware retention from the outset.

Safety guardrails have become an active area of research in AI safety, aimed at ensuring the appropriate behavior of large language models (LLMs). However, existing research lacks consideration of nuances across linguistic and cultural contexts, resulting in a gap between reported performance and in-the-wild effectiveness. To address this issue, this paper proposes an approach to optimize guardrail models for a designated linguistic context by leveraging a curated dataset tailored to local linguistic characteristics, targeting the Taiwan linguistic context as a representative example of localized deployment challenges. The proposed approach yields TWGuard, a linguistic context-optimized guardrail model that achieves a huge gain (+0.289 in F1) compared to the foundation model and significantly outperforms the strongest baseline in practical use (-0.037 in false positive rate, a 94.9% reduction). Together, this work lays a foundation for regional communities to establish AI safety standards grounded in their own linguistic contexts, rather than accepting boundaries imposed by dominant languages. The inadequacy of the latter is reconfirmed by our findings.

Mobile application developers are required to disclose how they collect, use, and share user data in compliance with privacy regulations. To support transparency, major app marketplaces have introduced standardized disclosure mechanisms. In 2022, Google mandated the Data Safety Section (DSS) on Google Play, requiring developers to summarize their data practices. However, compiling accurate DSS disclosures is challenging, as they must remain consistent with the corresponding privacy policy (PP), and no automated tool currently verifies this alignment. Prior studies indicate that nearly 80% of popular apps contain incomplete or misleading DSS declarations. We present PolicyGapper, an LLM-based methodology for automatically detecting discrepancies between DSS disclosures and privacy policies. PolicyGapper operates in four stages: scraping, pre-processing, analysis, and post-processing, without requiring access to application binaries. We evaluate PolicyGapper on a dataset of 330 top-ranked apps spanning all 33 Google Play categories, collected in Q3 2025. The approach identifies 2,689 omitted disclosures, including 2,040 related to data collection and 649 to data sharing. Manual validation on a stratified 10% subset, repeated across three independent runs, yields an average Precision of 0.75, Recall of 0.77, Accuracy of 0.69, and F1-score of 0.76. To support reproducibility, we release a complete replication package, including the dataset, prompts, source code, and results available at https://github.com/Mobile-IoT-Security-Lab/PolicyGapper and https://doi.org/10.5281/zenodo.19628493.

We propose a quantum-resistant quantum teleportation (QRQT) framework protected by post-quantum cryptography (PQC) to secure the classical correction channel, which is vulnerable to quantum adversaries. By applying PQC to the classical control bits, QRQT eliminates the classical attack surface of quantum teleportation. Our analysis reveals that quantum memory is a hidden bottleneck linking physical and computational security: its finite coherence time simultaneously limits communication distance, constrains tolerable PQC overhead, and restricts the adversary attack window. Under realistic parameters (1 ms coherence, fiber-optic propagation), the maximum secure teleportation distance ranges from 191 km (FrodoKEM-1344) to 199 km (Kyber512). We show that the joint classical-quantum attack probability exhibits a non-monotonic, Bell-shaped profile due to the opposing time dependencies of classical cryptanalysis and quantum decoherence, establishing a bounded optimal attack window beyond which adversarial success decays exponentially. We further analyze how leakage of classical correction bits affects teleportation security under four stochastic leakage models: independent exponential, sequential, burst, and correlated leakage, also accounting for amplitude damping on the shared Bell pair. For each scenario, we derive closed-form expressions for the average Holevo quantity and teleportation fidelity as functions of time, providing measurement-independent upper bounds on extractable information and guiding the design of leakage-resilient quantum communication protocols.

In most split-tunnel VPN/ZTNA deployments, installing an internal route authorizes the entire device, not a specific application, to use it. An unprivileged malicious process can therefore reach internal services by reusing routes intended for corporate applications. We present ProcRoute, a system that restricts internal-route access to explicitly authorized applications. ProcRoute models route access as an access-control problem: application identities are principals, destination prefixes with port and protocol constraints are resources, and a total, default-deny decision function mediates every connect() and UDP sendmsg() to an internal destination. Processes without a grant retain external access but are denied internal routes under our threat model. We describe ProcRoute's formal model, a Linux prototype built on cgroup v2 and eBPF socket-address hooks, and two complementary evaluations. In a two-machine WireGuard deployment, ProcRoute matches the WireGuard baseline and 13% faster than an nftables cgroup-matching configuration, with a p50 connect latency of 93 $\mu$s (+3.6 $\mu$s over baseline), flat policy scaling to 5,000 prefixes, and sub-millisecond revocation. Single-machine loopback microbenchmarks confirm low hook overhead: 2.7 $\mu$s on the internal-allow path, 82/82 unauthorized pivot attempts blocked, and zero transient allows across 1.2 million connection attempts during policy reload.

ISAC systems introduce new privacy risks because an unintended sensing node may exploit the shared radio waveform to infer transmitter-related information even when the communication payload remains secure. This paper investigates transmitter privacy, defined as limiting unauthorized inference of transmitter-related information through channel estimation, in a RIS-aided multi-antenna wireless system with a transmitter, a legitimate receiver, a malicious sensor, and a RIS. The malicious sensor is assumed to estimate the transmitter–sensor channel, and the resulting channel state information can then support unauthorized sensing, inference, or related signal processing. To mitigate this threat, we consider a privacy-oriented design in which the transmitter adopts superposition-based signaling with a message signal and transmit-side artificial noise, while the RIS shapes the propagation environment in a privacy-aware manner. The channel-estimation performance at the malicious sensor is first analyzed under imperfect prior knowledge, and both the true and predicted mean-square-error expressions are derived. Based on this analysis, we formulate a joint active–passive beamforming design problem that maximizes the malicious sensor's predicted channel-estimation error subject to a communication quality-of-service constraint, a transmit-power budget, and the unit-modulus constraints of the RIS. The resulting non-convex problem is handled through a numerically efficient alternating-optimization framework based on an augmented Lagrangian reformulation. Numerical results show that RIS-assisted propagation shaping can substantially degrade unauthorized channel estimation relative to the non-RIS case while preserving reliable communication, and further show that the privacy gains also improve a more direct sensing metric, namely the malicious sensor's angle-of-arrival estimation accuracy.

Understanding and anticipating vulnerability-related activity is a major challenge in cyber threat intelligence. This work investigates whether vulnerability sightings, such as proof-of-concept releases, detection templates, or online discussions, can be forecast over time. Building on our earlier work on VLAI, a transformer-based model that predicts vulnerability severity from textual descriptions, we examine whether severity scores can improve time-series forecasting as exogenous variables. We evaluate several approaches for short-term forecasting of sightings per vulnerability. First, we test SARIMAX models with and without log(x+1) transformations and VLAI-derived severity inputs. Although these adjustments provide limited improvements, SARIMAX remains poorly suited to sparse, short, and bursty vulnerability data. In practice, forecasts often produce overly wide confidence intervals and sometimes unrealistic negative values. To better capture the discrete and event-driven nature of sightings, we then explore count-based methods such as Poisson regression. Early results show that these models produce more stable and interpretable forecasts, especially when sightings are aggregated weekly. We also discuss simpler operational alternatives, including exponential decay functions for short forecasting horizons, to estimate future activity without requiring long historical series. Overall, this study highlights both the potential and the limitations of forecasting rare and bursty cyber events, and provides practical guidance for integrating predictive analytics into vulnerability intelligence workflows.

Code Large Language Models (Code LLMs) have revolutionized software development but raised critical concerns regarding code provenance, copyright protection, and security. Existing code watermarking approaches suffer from two fundamental limitations: black-box methods either exhibit detectable syntactic patterns vulnerable to statistical analysis or rely on implicit neural embedding behaviors that weaken interpretability, auditability, and precise control, while white-box methods lack code-aware capabilities that may compromise functionality. Moreover, current single-layer watermarking schemes fail to address increasingly complex provenance requirements such as multi-level attribution and version tracking. We present MATRIX, a novel code watermarking framework that formulates watermark encoding as solving constrained parity-check matrix equations. MATRIX employs dual-channel watermarking through variable naming and semantic-preserving transformations, enhancing watermark coverage across a wider range of code while ensuring mutual backup for robustness. By integrating BCH error-correction codes with solution space diversity, our approach achieves robustness against statistical analysis. Extensive evaluation on Python code generated by multiple Code LLMs demonstrates that MATRIX achieves an average watermark detection accuracy of 99.20% with minimal functionality loss (0-0.14%), improves robustness by 7.70-26.67% against various attacks, and increases watermarking applicability by 2-6x compared with existing methods. These results establish MATRIX as an effective solution for complex code provenance scenarios while balancing among detectability, fidelity, and robustness.

A central question of the Ethereum ecosystem is where Maximal Extractable Value (MEV)revenue originates and to what extent it stems from harming unsuspecting users. It is acceptable if MEV arises from arbitrages between centralised and decentralised exchanges (CEX-DEX). Yet theoretical models have significantly underestimated the scale of these arbitrages, while empirical studies have highlighted their importance - though these remain conservative estimates, constrained by numerous debatable heuristic assumptions. Revisiting the theoretical model, we found that CEX-DEX arbitrages require trading volumes on the order of the total activity of major liquidity pools and yield profits comparable to MEV. Most prior AMM models utilised the Black-Scholes (BS) stochastic differential equation (SDE) - i.e., geometric Brownian motion - and assumed continuous price trajectories where asset prices move in small increments only.We argue that BS underestimates arbitrage profits by ignoring price jumps, which are precisely the points at which arbitrage opportunities tend to arise. To address this gap, we present an extended discrete-time AMM model in which the price process is the sum of a diffusive component and stochastic jumps that can have arbitrary noise distributions. Although mathematically more involved this framework allows us to employ a general discrete-time SDE and compute the stationary probability distribution via function iteration with geometric convergence. We further prove that the resulting mispricing process is an ergodic Markov chain. We implement our model in C++, collect spot prices and AMM exchange data from the Ethereum blockchain and fit the model parameters to the observed prices. The estimates derived from our model closely match empirical observations and provide a natural theoretical explanation for several fundamental questions in the blockchain ecosystem.

Text-to-image (T2I) models are increasingly optimized for following user instructions faithfully. However, we find that this capability introduces a safety vulnerability we call Multi-Concept Compositional Unsafety (MCCU). MCCU occurs when multiple individually safe concepts, if combined in a single generation request, lead to harmful or sensitive visual outputs. Unlike prior jailbreak settings, MCCU does not rely on adversarial prompts, model access, or explicitly disallowed content. Instead, the risk emerges from how the model composes multiple safe visual concepts into a single scene. To systematically measure this threat, we build TwoHamsters, a large-scale evaluation framework consisting of 20k prompts, 51 curated concept pairs, and six risk categories. We evaluate 13 T2I models under a black-box setting. Our results show a clear conflict between instruction-following and safety: models that follow prompts more faithfully tend to produce more MCCU failures. For example, FLUX.1 achieves a 99.35% Unsafe Alignment Rate while only reaching a 1.57% MCCU Defense Rate. We further evaluate three representative defenses, including safety filtering, MCCU-specific detector fine-tuning, and concept erasure, all of which fail against unseen concept combinations. Our findings suggest that compositional reasoning in T2I models creates an attack surface that is not captured by existing safety mechanisms. We anticipate the release of TwoHamsters will catalyze community development of advanced generative defense mechanisms.

Despite the considerable promise of Retrieval-Augmented Generation (RAG), many real-world use cases may create privacy concerns, where the purported utility of RAG-enabled insights comes at the risk of exposing private information to either the LLM or the end user requesting the response. As a potential mitigation, using anonymization techniques to remove personally identifiable information (PII) and other sensitive markers in the underlying data represents a practical and sensible course of action for RAG administrators. Despite a wealth of literature on the topic, no works consider the placement of anonymization along the RAG pipeline, i.e., asking the question, where should anonymization happen? In this case study, we systematically and empirically measure the impact of anonymization at two important points along the RAG pipeline: the dataset and generated answer. We show that differences in privacy-utility trade-offs can be observed depending on where anonymization took place, demonstrating the significance of privacy risk mitigation placement in RAG.

We present a low-stack implementation of the module-lattice signature scheme HAETAE, targeting microcontrollers with 8 kB-16 kB of available SRAM. On such devices, peak stack usage is often the binding constraint, and HAETAE's hyperball-based sampler, large transient polynomial vectors, and variable-length signature payloads (hint and high-bits arrays) pose a particular challenge. To address this we introduce (i) Rejection-aware pass decomposition, which isolates encoding to the post-acceptance path; (ii) Component-level early rejection, which short-circuits the response computation when a partial norm already exceeds the bound; and (iii) Reverse-order streaming entropy coding using range Asymmetric Numeral Systems (rANS), which eliminates full hint and high-bits staging buffers. Combined with streamed matrix generation, a two-pass hyperball sampler with streaming Gaussian backend, and row-streamed verification, these techniques bring Signing stack from 71 kB-141 kB in the reference implementation down to 5.8 kB-6.0 kB, key generation to 4.7 kB-5.7 kB, and verification to 4.7 kB-4.8 kB across all three security levels. Our pure C implementation covers all three security levels (HAETAE-2/3/5), whose optimization paths differ due to the public-key domain (d>0 vs. d=0) and rejection structure. We implement our optimization on a Nucleo-L4R5ZI and compare to the reference pqm4 (for HAETAE-2 and -3) and a recently published memory-optimized implementation (targeting HAETAE-5 only). We reduce HAETAE-2, -3, and -5 stack by respectively 75, 86 and 8 % for key generation, 92, 95 and 24 % for signature generation, and 85, 91 and 22 % for verification. Depending on the parameter set, this impacts performance by at most a factor 1.8 and 3.4 for key and signature generation respectively, while even offering a performance improvement up to 18 % for verification. Verification at all security levels fits within 8 kB of RAM (signature buffer + stack) and is 2.34-3.34x faster than ML-DSA m4fstack at each comparable security level. We additionally validate portability under RIOT-OS on ARM Cortex-M4 and RISC-V targets.

Weber's conjecture (1886) governs three aspects of lattice-based cryptography: the solvability of the Principal Ideal Problem, the freeness of modules over rings of integers, and the tightness of worst-case-to-average-case reductions in Ring-LWE (R-LWE) and Module-LWE (MLWE). Existing verifications for $k \ge 9$ rely on Generalized Riemann Hypothesis (GRH). In this paper, we present the first unconditional proof for $k \le 12$. Our method combines the Fukuda-Komatsu computational sieve, inductive structure of the cyclotomic $\mathbb{Z}_2$-tower, and Herbrand's theorem.

Differential privacy (DP) has a wide range of applications for protecting data privacy, but designing and verifying DP algorithms requires expert-level reasoning, creating a high barrier for non-expert practitioners. Prior works either rely on specialized verification languages that demand substantial domain expertise or remain semi-automated and require human-in-the-loop guidance. In this work, we investigate whether large language models (LLMs) can automate DP reasoning. We introduce DPrivBench, a benchmark in which each instance asks whether a function or algorithm satisfies a stated DP guarantee under specified assumptions. The benchmark is carefully designed to cover a broad range of DP topics, span diverse difficulty levels, and resist shortcut reasoning through trivial pattern matching. Experiments show that while the strongest models handle textbook mechanisms well, all models struggle with advanced algorithms, revealing substantial gaps in current DP reasoning capabilities. Through further analytic study and failure-mode analysis, we identify several promising directions for improving automated DP reasoning. Our benchmark provides a solid foundation for developing and evaluating such methods, and complements existing benchmarks for mathematical reasoning.

Modern computing systems implicitly trust human input devices, allowing USB Human Interface Device (HID) emulators, such as the USB Rubber Ducky, to inject arbitrary keystrokes while bypassing conventional defenses. Speed- and regularity-based heuristics are easily evaded through slower or randomized timing. Keystroke dynamics offers a behavioral alternative, but prior work primarily addresses user authentication rather than user-agnostic detection of automated input. Moreover, detectors that process or retain key identities may expose sensitive content. In this paper, we investigate whether timing alone can distinguish human from machine-generated keystrokes without per-user enrollment or semantic input features. We systematically evaluate lightweight classifiers using only hold and flight times in a controlled offline setting. Across progressively structured synthetic attacker families, we examine cross-generator transfer, mixed-generator training, attacker sophistication, and the effect of observation-window length on discrimination performance and decision delay. Our results show that lightweight models can effectively discriminate human from machine-generated input across the evaluated generators without content access or user profiling. Our analysis also suggests that greater synthesis sophistication does not monotonically improve evasion. Moreover, detection performance primarily depends on exposure to structurally distinct generation strategies rather than model complexity. This finding can reduce the number of attacker models required for training and supports the feasibility of content-independent, user-agnostic HID-injection detection, while physical HID validation remains necessary.