Papers

Cross-border insider threats pose a critical challenge to government financial schemes, particularly when dealing with distributed, privacy-sensitive data across multiple jurisdictions. Existing approaches face fundamental limitations: they cannot effectively share intelligence across borders due to privacy constraints, lack reasoning capabilities to understand complex multi-step attack patterns, and fail to capture intricate graph-structured relationships in financial networks. We introduce FedGraph-AGI, a novel federated learning framework integrating Artificial General Intelligence (AGI) reasoning with graph neural networks for privacy-preserving cross-border insider threat detection. Our approach combines: (1) federated graph neural networks preserving data sovereignty; (2) Mixture-of-Experts (MoE) aggregation for heterogeneous jurisdictions; and (3) AGI-powered reasoning via Large Action Models (LAM) performing causal inference over graph data. Through experiments on a 50,000-transaction dataset across 10 jurisdictions, FedGraph-AGI achieves 92.3% accuracy, significantly outperforming federated baselines (86.1%) and centralized approaches (84.7%). Our ablation studies reveal AGI reasoning contributes 6.8% improvement, while MoE adds 4.4%. The system maintains epsilon = 1.0 differential privacy while achieving near-optimal performance and scales efficiently to 50+ clients. This represents the first integration of AGI reasoning with federated graph learning for insider threat detection, opening new directions for privacy-preserving cross-border intelligence sharing.

The Internet of Battlefield Things (IoBT) relies on heterogeneous, bandwidth-constrained, and intermittently connected tactical networks that face rapidly evolving cyber threats. In this setting, intrusion detection cannot depend on continuous central collection of raw traffic due to disrupted links, latency, operational security limits, and non-IID traffic across zones. We present Zone-Adaptive Intrusion Detection (ZAID), a collaborative detection and model-improvement framework for unseen attack types, where "zero-day" refers to previously unobserved attack families and behaviours (not vulnerability disclosure timing). ZAID combines a universal convolutional model for generalisable traffic representations, an autoencoder-based reconstruction signal as an auxiliary anomaly score, and lightweight adapter modules for parameter-efficient zone adaptation. To support cross-zone generalisation under constrained connectivity, ZAID uses federated aggregation and pseudo-labelling to leverage locally observed, weakly labelled behaviours. We evaluate ZAID on ToN_IoT using a zero-day protocol that excludes MITM, DDoS, and DoS from supervised training and introduces them during zone-level deployment and adaptation. ZAID achieves up to 83.16% accuracy on unseen attack traffic and transfers to UNSW-NB15 under the same procedure, with a best accuracy of 71.64%. These results indicate that parameter-efficient, zone-personalised collaboration can improve the detection of previously unseen attacks in contested IoBT environments.

High-frequency banking environments face a critical trade-off between low-latency fraud detection and the regulatory explainability demanded by GDPR. Traditional rule-based and discriminative models struggle with "zero-day" attacks due to extreme class imbalance and the lack of historical precedents. This paper proposes a Dual-Path Generative Framework that decouples real-time anomaly detection from offline adversarial training. The architecture employs a Variational Autoencoder (VAE) to establish a legitimate transaction manifold based on reconstruction error, ensuring <50ms inference latency. In parallel, an asynchronous Wasserstein GAN with Gradient Penalty (WGAN-GP) synthesizes high-entropy fraudulent scenarios to stress-test the detection boundaries. Crucially, to address the non-differentiability of discrete banking data (e.g., Merchant Category Codes), we integrate a Gumbel-Softmax estimator. Furthermore, we introduce a trigger-based explainability mechanism where SHAP (Shapley Additive Explanations) is activated only for high-uncertainty transactions, reconciling the computational cost of XAI with real-time throughput requirements.

We define a method how digital ecosystems (including data spaces) may autonomously define and "advertise" credentials they issue or they trust in the form of so-called ecosystem trust profiles. An ecosystem trust profile collects all (verifiable) credentials and issuers sorted by trust scope accepted ("trusted") by a particular ecosystem. We then show how a minimal trust relation between ecosystems may be defined using ecosystem trust frameworks of different ecosystems and explore a few of its properties. A first application of the theory is given for a use case in the manufacturing realm where different international ecosystems need to agree on certain credentials for various scopes of trust such as identity, service compliance, and other conformance standards. We implement this requirement by identifying and discussing two different definitions of credential equivalence for a given trust scope, one requiring additional cross-ecosystem governance or coordination, one not. The second approach demonstrates how to solve the so-called cross-ecosystem trust dilemma, that is, the problem how ecosystems can establish cross-ecosystem trust while, at the same time, allowing them to fully retain their sovereignty. A fragility theorem demonstrates that this sovereignty leads trust to be unstable without any additional coordination or governance mechanisms on top of (and outside to) ecosystem trust profiles. We extend our method to data spaces in particular and propose a novel rigorous definition of cross-data space interoperability. This allows us to prove the proposition that the extent of interoperability between two data spaces is exactly determined by the amount of commonality in their respective ecosystem trust profiles.

We present SuperLocalMemory, a local-first memory system for multi-agent AI that defends against OWASP ASI06 memory poisoning through architectural isolation and Bayesian trust scoring, while personalizing retrieval through adaptive learning-to-rank – all without cloud dependencies or LLM inference calls. As AI agents increasingly rely on persistent memory, cloud-based memory systems create centralized attack surfaces where poisoned memories propagate across sessions and users – a threat demonstrated in documented attacks against production systems. Our architecture combines SQLite-backed storage with FTS5 full-text search, Leiden-based knowledge graph clustering, an event-driven coordination layer with per-agent provenance, and an adaptive re-ranking framework that learns user preferences through three-layer behavioral analysis (cross-project technology preferences, project context detection, and workflow pattern mining). Evaluation across seven benchmark dimensions demonstrates 10.6ms median search latency, zero concurrency errors under 10 simultaneous agents, trust separation (gap =0.90) with 72% trust degradation for sleeper attacks, and 104% improvement in NDCG@5 when adaptive re-ranking is enabled. Behavioral data is isolated in a separate database with GDPR Article 17 erasure support. SuperLocalMemory is open-source (MIT) and integrates with 17+ development tools via Model Context Protocol.

We systematically evaluate the quality of widely used adversarial safety datasets from two perspectives: in isolation and in practice. In isolation, we examine how well these datasets reflect real-world adversarial attacks based on three defining properties: being driven by ulterior intent, well-crafted, and out-of-distribution. We find that these datasets overrely on "triggering cues": words or phrases with overt negative/sensitive connotations that are intended to trigger safety mechanisms explicitly, which is unrealistic compared to real-world attacks. In practice, we evaluate whether these datasets genuinely measure safety risks or merely provoke refusals through triggering cues. To explore this, we introduce "intent laundering": a procedure that abstracts away triggering cues from adversarial attacks (data points) while strictly preserving their malicious intent and all relevant details. Our results show that current adversarial safety datasets fail to faithfully represent real-world adversarial behavior due to their overreliance on triggering cues. Once these cues are removed, all previously evaluated "reasonably safe" models become unsafe, including Gemini 3 Pro and Claude Sonnet 3.7/4. Moreover, when intent laundering is adapted as a jailbreaking technique, it consistently achieves high attack success rates, ranging from 90.00% to 100.00%, under fully black-box access. Overall, our findings expose a significant disconnect between how existing datasets evaluate model safety and how real-world adversaries behave.

Analog processing-using-memory (PUM; a.k.a. in-memory computing) makes use of electrical interactions inside memory arrays to perform bulk matrix-vector multiplication (MVM) operations. However, many popular matrix-based kernels need to execute non-MVM operations, which analog PUM cannot directly perform. To retain its energy efficiency, analog PUM architectures augment memory arrays with CMOS-based domain-specific fixed-function hardware to provide complete kernel functionality, but the difficulty of integrating such specialized CMOS logic with memory arrays has largely limited analog PUM to being an accelerator for machine learning inference, or for closely related kernels. An opportunity exists to harness analog PUM for general-purpose computation: recent works have shown that memory arrays can also perform Boolean PUM operations, albeit with very different supporting hardware and electrical signals than analog PUM. We propose DARTH-PUM, a general-purpose hybrid PUM architecture that tackles key hardware and software challenges to integrating analog PUM and digital PUM. We propose optimized peripheral circuitry, coordinating hardware to manage and interface between both types of PUM, an easy-to-use programming interface, and low-cost support for flexible data widths. These design elements allow us to build a practical PUM architecture that can execute kernels fully in memory, and can scale easily to cater to domains ranging from embedded applications to large-scale data-driven computing. We show how three popular applications (AES encryption, convolutional neural networks, large language models) can map to and benefit from DARTH-PUM, with speedups of 59.4x, 14.8x, and 40.8x over an analog+CPU baseline.

In this paper, we explore the short- and long-term stability of backed stablecoins offering constant mint and redeem prices to all agents. We refer to such designs as price window-based, since the mint and redeem prices constrain the stablecoin's market equilibrium. We show that, without secondary stabilization mechanisms, price window designs cannot achieve both short- and long-term stability unless they are backed by already-stable reserves. In particular, the mechanism faces a tradeoff: either risk eventual reserve depletion through persistent arbitrage by a speculator, or widen the distance between mint and redeem prices enough to disincentivize arbitrage. In the latter case, however, the market price of the stablecoin inherits the volatility of its backing asset, with fluctuations that can be proportional to the backing asset's own volatility.

As cyber threats grow in complexity and scale, many security incidents remain poorly managed due to the lack of proper training among C-level executives. Thus, there is a need for targeted cybersecurity education to enhance executive decision-making and crisis response. Traditional training methods, such as cyber wargames and Tabletop Exercises (TTX), aim to develop abilities to face critical incidents, however, they often lack the interactive and dynamic elements required to prepare individuals for real-world cyber incidents. This paper presents a novel approach to cybersecurity and cyberdefense education through the design of a specialized hybrid TTX for the maritime domain, which uses a framework to model mathematically how a cyberattack spreads along multiple nodes and impacts infrastructure. Our proposal was validated through exercises in Argentina and the United States, demonstrating a positive impact in developing the comprehension and projection levels of Cyber Situational Awareness (CSA), and reinforcing governance. Documentation about the Hybrid TTX, scenario, datasets and implementation of the SERDUX-MARCIM model, is available at the project repository at https://github.com/diegocabuya/SERDUX-MARCIM

We study a sequential coherent side-channel model in which an adversarial probe qubit interacts with a target qubit during a hidden gate sequence. Repeating the same hidden sequence for $N$ shots yields an empirical full-correlation record: the joint histogram $\widehat{P}_g(b)$ over probe bit-strings $b\in\{0,1\}^k$, which is a sufficient statistic for classical post-processing under identically and independently distributed (i.i.d.)\ shots but grows exponentially with circuit depth. We first describe this sequential probe framework in a coupling- and measurement-agnostic form, emphasizing the scaling of the observation space and why exact analytic distinguishability becomes intractable with circuit depth. We then specialize to a representative instantiation (a controlled-rotation probe coupling with fixed projective readout and a commuting $R_x$ gate alphabet) where we (i) derive a depth-dependent leakage envelope whose maximizer predicts a coupling band as a function of depth {if the measurement data is reduced to marginal statistics}, and (ii) provide an operational decoder, via machine learning, a single parameter-conditioned map from $\widehat{P}_g$ to Alice's per-step gate labels, generalizing across coupling and noise settings without retraining.

Model Context Protocols (MCPs) provide a unified platform for agent systems to discover, select, and orchestrate tools across heterogeneous execution environments. As MCP-based systems scale to incorporate larger tool catalogs and multiple concurrently connected MCP servers, traditional tool-by-tool invocation increases coordination overhead, fragments state management, and limits support for wide-context operations. To address these scalability challenges, recent MCP designs have incorporated code execution as a first-class capability, an approach called Code Execution MCP (CE-MCP). This enables agents to consolidate complex workflows, such as SQL querying, file analysis, and multi-step data transformations, into a single program that executes within an isolated runtime environment. In this work, we formalize the architectural distinction between context-coupled (traditional) and context-decoupled (CE-MCP) models, analyzing their fundamental scalability trade-offs. Using the MCP-Bench framework across 10 representative servers, we empirically evaluate task behavior, tool utilization patterns, execution latency, and protocol efficiency as the scale of connected MCP servers and available tools increases, demonstrating that while CE-MCP significantly reduces token usage and execution latency, it introduces a vastly expanded attack surface. We address this security gap by applying the MAESTRO framework, identifying sixteen attack classes across five execution phases-including specific code execution threats such as exception-mediated code injection and unsafe capability synthesis. We validate these vulnerabilities through adversarial scenarios across multiple LLMs and propose a layered defense architecture comprising containerized sandboxing and semantic gating. Our findings provide a rigorous roadmap for balancing scalability and security in production-ready executable agent workflows.

We study privacy filters, which enable privacy accounting for differentially private (DP) mechanisms with adaptively chosen privacy characteristics. We develop a general theory that characterizes the worst-case privacy loss of an interaction involving an analyst that respects some restrictions on what queries they may issue. We apply this theory to develop residue filters, which unifies existing privacy filters. We develop the Gaussian DP (GDP) residue filter, which strictly improves upon the naïve GDP filter. We also show that residue filters capture the natural filter, which promises greater utility by leveraging exact privacy accounting techniques. Earlier privacy filters consider only simple privacy parameters such as Rényi-DP or GDP parameters. Natural filters account for the entire privacy profile of every query, promising more efficient use of a given privacy budget. We show that, contrary to other forms of DP, natural privacy filters are not free in general. We present a characterization of when a family of private queries admits free natural filters for a given budget. In particular, only families of privacy mechanisms that are totally-ordered when composed admit free natural privacy filters with respect to an arbitrary privacy budget. Finally, we show that, while the natural approximate-DP filter can fail in the presence of adaptive adversary, it cannot fail too badly: the output remains approximate-DP with parameters at most poly-logarithmically worse than the intended privacy parameters.

We initiate an investigation of node differential privacy for graphs in the local model of private data analysis. In our model, dubbed LNDP*, each node sees its own edge list and releases the output of a local randomizer on this input. These outputs are aggregated by an untrusted server to obtain a final output. We develop a novel algorithmic framework for this setting that allows us to accurately answer arbitrary linear queries about the input graph's degree distribution. Our framework is based on a new object, called the blurry degree distribution, which closely approximates the degree distribution and has lower sensitivity. Instead of answering queries about the degree distribution directly, our algorithms answer queries about the blurry degree distribution. This framework yields accurate LNDP* algorithms for the edge count, PMF and CDF of the degree distribution, and other graph statistics. For some natural problems, our algorithms match the accuracy achievable with node privacy in the central model, where data are held and processed by a trusted server. We also prove lower bounds on the error required by LNDP* algorithms that imply the optimality of our framework for edge counting in sparse graphs and Erdos-Renyi parameter estimation. Our lower bounds apply even to interactive protocols with a constant number of rounds of interaction between the nodes and the server. Existing lower-bound techniques for related models either yield loose bounds or do not apply in our setting, as graph data results in inherently overlapping inputs to local randomizers. To prove our bounds, we develop a splicing argument that stitches together views from locally similar but globally different distributions on graphs to obtain hard instances. Finally, we prove structural results that reveal qualitative differences between local node privacy and the standard local model for tabular data.

A natural and informal approach to verifiable (or zero-knowledge) ML inference over floating-point data is: “prove that each layer was computed correctly up to tolerance $\delta$; therefore the final output is a reasonable inference result”. This short note gives a simple counterexample showing that this inference is false in general: for any neural network, we can construct a functionally equivalent network for which adversarially chosen approximation-magnitude errors in individual layer computations suffice to steer the final output arbitrarily (within a prescribed bounded range).

Database-driven cognitive radio networks (DB-CRNs) enable dynamic spectrum sharing through geolocation databases but introduce critical security and privacy challenges, including mandatory location disclosure, susceptibility to location spoofing, and denial-of-service (DoS) attacks on centralized services. Existing approaches address these issues in isolation and lack a unified, regulation-compliant solution under realistic adversarial conditions. In this work, we present a unified security framework for DB-CRNs that simultaneously provides location privacy, user anonymity, verifiable location, and DoS resilience. Our framework, denoted as SLAPX, enables privacy-preserving spectrum queries using delegatable anonymous credentials, supports adaptive location verification without revealing precise user location, and mitigates DoS attacks through verifiable delay functions (VDFs) combined with RLRS-based rate limiting. Extensive cryptographic benchmarking and network simulations demonstrate that SLAPX achieves significantly lower latency and communication overhead than existing solutions while effectively resisting location spoofing and DoS attacks. These results show that SLAPX is practical and well-suited for secure next-generation DB-CRN deployments.

Large language models and LLM-based agents are increasingly used for cybersecurity tasks that are inherently dual-use. Existing approaches to refusal, spanning academic policy frameworks and commercially deployed systems, often rely on broad topic-based bans or offensive-focused taxonomies. As a result, they can yield inconsistent decisions, over-restrict legitimate defenders, and behave brittlely under obfuscation or request segmentation. We argue that effective refusal requires explicitly modeling the trade-off between offensive risk and defensive benefit, rather than relying solely on intent or offensive classification. In this paper, we introduce a content-based framework for designing and auditing cyber refusal policies that makes offense-defense tradeoffs explicit. The framework characterizes requests along five dimensions: Offensive Action Contribution, Offensive Risk, Technical Complexity, Defensive Benefit, and Expected Frequency for Legitimate Users, grounded in the technical substance of the request rather than stated intent. We demonstrate that this content-grounded approach resolves inconsistencies in current frontier model behavior and allows organizations to construct tunable, risk-aware refusal policies.

Federated learning security research has predominantly focused on backdoor threats from a minority of malicious clients that intentionally corrupt model updates. This paper challenges this paradigm by investigating a more pervasive and insidious threat: backdoor vulnerabilities from low-concentration poisoned data distributed across the datasets of benign clients. This scenario is increasingly common in federated instruction tuning for language models, which often rely on unverified third-party and crowd-sourced data. We analyze two forms of backdoor data through real cases: 1) natural trigger (inherent features as implicit triggers); 2) adversary-injected trigger. To analyze this threat, we model the backdoor implantation process from signal aggregation, proposing the Backdoor Signal-to-Noise Ratio to quantify the dynamics of the distributed backdoor signal. Extensive experiments reveal the severity of this threat: With just less than 10% of training data poisoned and distributed across clients, the attack success rate exceeds 85%, while the primary task performance remains largely intact. Critically, we demonstrate that state-of-the-art backdoor defenses, designed for attacks from malicious clients, are fundamentally ineffective against this threat. Our findings highlight an urgent need for new defense mechanisms tailored to the realities of modern, decentralized data ecosystems.

Self-evolving LLM agents update their internal state across sessions, often by writing and reusing long-term memory. This design improves performance on long-horizon tasks but creates a security risk: untrusted external content observed during a benign session can be stored as memory and later treated as instruction. We study this risk and formalize a persistent attack we call a Zombie Agent, where an attacker covertly implants a payload that survives across sessions, effectively turning the agent into a puppet of the attacker. We present a black-box attack framework that uses only indirect exposure through attacker-controlled web content. The attack has two phases. During infection, the agent reads a poisoned source while completing a benign task and writes the payload into long-term memory through its normal update process. During trigger, the payload is retrieved or carried forward and causes unauthorized tool behavior. We design mechanism-specific persistence strategies for common memory implementations, including sliding-window and retrieval-augmented memory, to resist truncation and relevance filtering. We evaluate the attack on representative agent setups and tasks, measuring both persistence over time and the ability to induce unauthorized actions while preserving benign task quality. Our results show that memory evolution can convert one-time indirect injection into persistent compromise, which suggests that defenses focused only on per-session prompt filtering are not sufficient for self-evolving agents.

In this paper, we investigate how attackers can discover sensitive information embedded within databases by exploiting inference rules. We demonstrate the inadequacy of naively applied existing state of the art differential privacy (DP) models in safeguarding against such attacks. We introduce ontology aware differential privacy (Onto-DP), a novel extension of differential privacy paradigms built on top of any classical DP model by enriching it with semantic awareness. We show that this extension is a sufficient condition to adequately protect against attackers aware of inference rules.

We introduce SecCodeBench-V2, a publicly released benchmark for evaluating Large Language Model (LLM) copilots' capabilities of generating secure code. SecCodeBench-V2 comprises 98 generation and fix scenarios derived from Alibaba Group's industrial productions, where the underlying security issues span 22 common CWE (Common Weakness Enumeration) categories across five programming languages: Java, C, Python, Go, and JavaScript. SecCodeBench-V2 adopts a function-level task formulation: each scenario provides a complete project scaffold and requires the model to implement or patch a designated target function under fixed interfaces and dependencies. For each scenario, SecCodeBench-V2 provides executable proof-of-concept (PoC) test cases for both functional validation and security verification. All test cases are authored and double-reviewed by security experts, ensuring high fidelity, broad coverage, and reliable ground truth. Beyond the benchmark itself, we build a unified evaluation pipeline that assesses models primarily via dynamic execution. For most scenarios, we compile and run model-generated artifacts in isolated environments and execute PoC test cases to validate both functional correctness and security properties. For scenarios where security issues cannot be adjudicated with deterministic test cases, we additionally employ an LLM-as-a-judge oracle. To summarize performance across heterogeneous scenarios and difficulty levels, we design a Pass@K-based scoring protocol with principled aggregation over scenarios and severity, enabling holistic and comparable evaluation across models. Overall, SecCodeBench-V2 provides a rigorous and reproducible foundation for assessing the security posture of AI coding assistants, with results and artifacts released at https://alibaba.github.io/sec-code-bench. The benchmark is publicly available at https://github.com/alibaba/sec-code-bench.

We study builder-driven MEV arbitrage on BNB Smart Chain (BSC). BSC's Proposer-Builder Separation (PBS) adopts a leaner design: only whitelisted builders can participate, blocks are produced at shorter intervals, and private order flow bypasses the public mempool. These features have long raised community concerns over centralization, which we empirically confirm by tracing the arbitrage activities of the two dominant builders from Apr. 1, 2025 to Feb. 28, 2026 (full observable activity cycle). Within months, the two leading builders, \bd{48Club} and \bd{Blockrazor}, produced over 87% of blocks and captured about 90%+ of MEV profits. We find that profits concentrate in short, low-hop arbitrage routes over wrapped tokens and stablecoins, and that block construction rapidly converges toward monopoly. Beyond concentration alone, our analysis reveals a structural source of inequality: BSC's short block interval and whitelisted PBS collapse the contestable window for MEV competition, amplifying latency advantages and excluding slower builders and searchers. MEV extraction on BSC is not only more centralized than on Ethereum, but also structurally more vulnerable to censorship and fairness erosion.

Cryptographic digests (e.g., MD5, SHA-256) are designed to provide exact identity. Any single-bit change in the input produces a completely different hash, which is ideal for integrity verification but limits their usefulness in many real-world tasks like threat hunting, malware analysis and digital forensics, where adversaries routinely introduce minor transformations. Similarity-based techniques address this limitation by enabling approximate matching, allowing related byte sequences to produce measurably similar fingerprints. Modern enterprises manage tens of thousands of endpoints with billions of files, making the effectiveness and scalability of the proposed techniques more important than ever in security applications. Security researchers have proposed a range of approaches, including similarity digests and locality-sensitive hashes (e.g., ssdeep, sdhash, TLSH), as well as more recent machine-learning-based methods that generate embeddings from file features. However, these techniques have largely been evaluated in isolation, using disparate datasets and evaluation criteria. This paper presents a systematic comparison of learning-based classification and similarity methods using large, publicly available datasets. We evaluate each method under a unified experimental framework with industry-accepted metrics. To our knowledge, this is the first reproducible study to benchmark these diverse learning-based similarity techniques side by side for real-world security workloads. Our results show that no single approach performs well across all dimensions; instead, each exhibits distinct trade-offs, indicating that effective malware analysis and threat-hunting platforms must combine complementary classification and similarity techniques rather than rely on a single method.

AI watermarking embeds invisible signals within images to provide provenance information and identify content as AI-generated. In this paper, we introduce MarkSweep, a novel watermark removal attack that effectively erases the embedded watermarks from AI-generated images without degrading visual quality. MarkSweep first amplifies watermark noise in high-frequency regions via edge-aware Gaussian perturbations and injects it into clean images for training a denoising network. This network then integrates two modules, the learnable frequency decomposition module and the frequency-aware fusion module, to suppress amplified noise and eliminate watermark traces. Theoretical analysis and extensive experiments demonstrate that invisible watermarks are highly vulnerable to MarkSweep, which effectively removes embedded watermarks, reducing the bit accuracy of HiDDeN and Stable Signature watermarking schemes to below 67%, while preserving perceptual quality of AI-generated images.

Language models now routinely produce text that is difficult to distinguish from human writing, raising the need for robust tools to verify content provenance. Watermarking has emerged as a promising countermeasure, with existing work largely focused on model quality preservation and robust detection. However, current schemes provide limited protection against false attribution. We strengthen the notion of soundness by introducing two novel guarantees: unforgeability and recoverability. Unforgeability prevents adversaries from crafting false positives, texts that are far from any output from the watermarked model but are nonetheless flagged as watermarked. Recoverability provides an additional layer of protection: whenever a watermark is detected, the detector identifies the source text from which the flagged content was derived. Together, these properties strengthen content ownership by linking content exclusively to its generating model, enabling secure attribution and fine-grained traceability. We construct the first undetectable watermarking scheme that is robust, unforgeable, and recoverable with respect to substitutions (i.e., perturbations in Hamming metric). The key technical ingredient is a new cryptographic primitive called robust (or recoverable) digital signatures, which allow verification of messages that are close to signed ones, while preventing forgery of messages that are far from all previously signed messages. We show that any standard digital signature scheme can be boosted to a robust one using property-preserving hash functions (Boyle, LaVigne, and Vaikuntanathan, ITCS 2019).

Intellicise (Intelligent and Concise) wireless network is the main direction of the evolution of future mobile communication systems, a perspective now widely acknowledged across academia and industry. As a key technology within it, Agentic AI has garnered growing attention due to its advanced cognitive capabilities, enabled through continuous perception-memory-reasoning-action cycles. This paper first analyses the unique advantages that Agentic AI introduces to intellicise wireless networks. We then propose a structured taxonomy for Agentic AI-enhanced secure intellicise wireless networks. Building on this framework, we identify emerging security and privacy challenges introduced by Agentic AI and summarize targeted strategies to address these vulnerabilities. A case study further demonstrates Agentic AI's efficacy in defending against intelligent eavesdropping attacks. Finally, we outline key open research directions to guide future exploration in this field.

State-space models like Mamba offer linear-time sequence processing and low memory, making them attractive for medical imaging. However, their robustness under realistic software and hardware threat models remains underexplored. This paper evaluates Mamba on multiple MedM-NIST classification benchmarks under input-level attacks, including white-box adversarial perturbations (FGSM/PGD), occlusion-based PatchDrop, and common acquisition corruptions (Gaussian noise and defocus blur) as well as hardware-inspired fault attacks emulated in software via targeted and random bit-flip injections into weights and activations. We profile vulnerabilities and quantify impacts on accuracy indicating that defenses are needed for deployment.

An open measurement problem in IoT security is whether scan-observable network configurations encode population-level exposure risk beyond individual devices. An analysis of internet-exposed IoT endpoints using a controlled multi-country sample from Shodan Search and Shodan InternetDB, selecting 100 hosts identified via TCP port 7547 (TR-069/CWMP) and evenly distributed across the ten most represented countries. Hosts are enriched with scan-derived metadata and analyzed using feature-relevance assessment, cross-country comparisons of open and risky port exposure, and supervised classification of higher-risk exposure profiles. The analysis reveals consistent cross-country differences in exposure structure, with mean risky-port counts ranging from 0.4 to 1.0 per host, and achieves balanced accuracy of approximately 0.61 when classifying higher-risk exposure profiles.

Adversarial training for LLMs is one of the most promising methods to reliably improve robustness against adversaries. However, despite significant progress, models remain vulnerable to simple in-distribution exploits, such as rewriting prompts in the past tense or translating them into other languages. We argue that this persistent fragility stems from a fundamental limitation in current adversarial training algorithms: they minimize adversarial loss on their training set but inadequately cover the data distribution, resulting in vulnerability to seemingly simple attacks. To bridge this gap, we propose Distributional Adversarial Training, DAT. We leverage Diffusion LLMs to approximate the true joint distribution of prompts and responses, enabling generation of diverse, high-likelihood samples that address generalization failures. By combining optimization over the data distribution provided by the diffusion model with continuous adversarial training, DAT achieves substantially higher adversarial robustness than previous methods.

LoRA adapters let users fine-tune large language models (LLMs) efficiently. However, LoRA adapters are shared through open repositories like Hugging Face Hub \citep{huggingface_hub_docs}, making them vulnerable to backdoor attacks. Current detection methods require running the model with test input data – making them impractical for screening thousands of adapters where the trigger for backdoor behavior is unknown. We detect poisoned adapters by analyzing their weight matrices directly, without running the model – making our method trigger-agnostic. For each attention projection (Q, K, V, O), our method extracts five spectral statistics from the low-rank update $\Delta W$, yielding a 20-dimensional signature for each adapter. A logistic regression detector trained on this representation separates benign and poisoned adapters across three model families – Llama-3.2-3B~\citep{llama3}, Qwen2.5-3B~\citep{qwen25}, and Gemma-2-2B~\citep{gemma2} – on unseen test adapters drawn from instruction-following, reasoning, question-answering, code, and classification tasks. Across all three architectures, the detector achieves 100% accuracy.

Federated learning (FL) enables distributed model training across edge devices while preserving data locality. This decentralized approach has emerged as a promising solution for collaborative learning on sensitive user data, effectively addressing the longstanding privacy concerns inherent in centralized systems. However, the decentralized nature of FL exposes new security vulnerabilities, especially backdoor attacks that threaten model integrity. To investigate this critical concern, this paper presents the Layer Smoothing Attack (LSA), a novel backdoor attack that exploits layer-specific vulnerabilities in neural networks. First, a Layer Substitution Analysis methodology systematically identifies backdoor-critical (BC) layers that contribute most significantly to backdoor success. Subsequently, LSA strategically manipulates these BC layers to inject persistent backdoors while remaining undetected by state-of-the-art defense mechanisms. Extensive experiments across diverse model architectures and datasets demonstrate that LSA achieves a remarkably backdoor success rate of up to 97% while maintaining high model accuracy on the primary task, consistently bypassing modern FL defenses. These findings uncover fundamental vulnerabilities in current FL security frameworks, demonstrating that future defenses must incorporate layer-aware detection and mitigation strategies.