Papers

Federated Learning (FL) facilitates collaborative model training among distributed clients while ensuring that raw data remains on local devices.Despite this advantage, FL systems are still exposed to risks from malicious or unreliable participants. Such clients can interfere with the training process by sending misleading updates, which can negatively affect the performance and reliability of the global model. Many existing defense mechanisms rely on gradient inspection, complex similarity computations, or cryptographic operations, which introduce additional overhead and may become unstable under non-IID data distributions. In this paper, we propose the Federated Learning with Loss Trend Detection (FL-LTD), a lightweight and privacy-preserving defense framework that detects and mitigates malicious behavior by monitoring temporal loss dynamics rather than model gradients. The proposed approach identifies anomalous clients by detecting abnormal loss stagnation or abrupt loss fluctuations across communication rounds. To counter adaptive attackers, a short-term memory mechanism is incorporated to sustain mitigation for clients previously flagged as anomalous, while enabling trust recovery for stable participants. We evaluate FL-LTD on a non-IID federated MNIST setup under loss manipulation attacks. Experimental results demonstrate that the proposed method significantly enhances robustness, achieving a final test accuracy of 0.84, compared to 0.41 for standard FedAvg under attack. FL-LTD incurs negligible computational and communication overhead, maintains stable convergence, and avoids client exclusion or access to sensitive data, highlighting the effectiveness of loss-based monitoring for secure federated learning.

Multi-turn jailbreak attacks have emerged as a critical threat to Large Language Models (LLMs), bypassing safety mechanisms by progressively constructing adversarial contexts from scratch and incrementally refining prompts. However, existing methods suffer from the inefficiency of incremental context construction that requires step-by-step LLM interaction, and often stagnate in suboptimal regions due to surface-level optimization. In this paper, we characterize the Intent-Context Coupling phenomenon, revealing that LLM safety constraints are significantly relaxed when a malicious intent is coupled with a semantically congruent context pattern. Driven by this insight, we propose ICON, an automated multi-turn jailbreak framework that efficiently constructs an authoritative-style context via prior-guided semantic routing. Specifically, ICON first routes the malicious intent to a congruent context pattern (e.g., Scientific Research) and instantiates it into an attack prompt sequence. This sequence progressively builds the authoritative-style context and ultimately elicits prohibited content. In addition, ICON incorporates a Hierarchical Optimization Strategy that combines local prompt refinement with global context switching, preventing the attack from stagnating in ineffective contexts. Experimental results across eight SOTA LLMs demonstrate the effectiveness of ICON, achieving a state-of-the-art average Attack Success Rate (ASR) of 97.1%. Code is available at https://github.com/xwlin-roy/ICON.

With LLMs increasingly deployed in corporate data management, it is crucial to ensure that these models do not leak sensitive information. In the context of corporate data management, the concept of sensitivity awareness has been introduced, enabling LLMs to adhere to predefined access rights rules. However, it remains unclear how sensitivity awareness relates to established notions of privacy, such as differential privacy (DP), thereby making it difficult to deploy meaningfully in real-world applications. In this work, we formalize the notion of sensitivity awareness and theoretically establish its connection to DP. Additionally, we develop a supervised fine-tuning recipe to make existing, four-bit quantized LLMs more sensitivity-aware. With a performance boost of up to 21.7%, the finetuned LLMs not only substantially improve over their baseline but also outperform other full-precision open-source and commercial models of similar size in achieving sensitivity awareness, demonstrating the effectiveness of our proposed approach. At the same time, our method also largely preserves the models' performance on other tasks, such as general instruction-following, mathematical, and common-sense reasoning.

Decentralized Identifiers (DIDs) are increasingly deployed on distributed ledgers, yet systematic cross-platform evidence on their operational behavior remains limited. We present an empirical benchmarking study of three prominent ledger-based DID methods - Ethereum, Hedera, and XRP Ledger - using reference Software Development Kits (SDKs) under a unified experimental setup. We measure latency, transaction cost, and on-chain metadata exposure, normalizing latency by each platform's block or consensus interval and cost by its native value transfer fee. Privacy leakage is quantified using a Metadata-Leakage Score (MLS), an entropy-based measure expressed in bits per operation. Our results reveal distinct architectural trade-offs. Ethereum enables near-instant, off-chain DID creation, but incurs the highest latency and cost for on-chain lifecycle operations. XRPL delivers deterministic and stable latency with fixed, low fees, yet exhibits higher metadata leakage due to more verbose transaction payloads. Hedera achieves the lowest on-chain latency and low fees with minimal metadata leakage, while occasional variance arises from SDK-side processing and confirmation pipelines. Overall, the findings show that ledger architecture and SDK workflows play a major role in shaping DID latency, cost, and metadata exposure, complementing the effects of the underlying consensus mechanism. These results provide evidence-based insights to support informed selection and configuration of DID systems under performance and privacy constraints.

Dependency management systems are a critical component in software development, enabling projects to incorporate existing functionality efficiently. However, misconfigurations and malicious actors in these systems pose severe security risks, leading to supply chain attacks. Despite the widespread use of smartphone apps, the security of dependency management systems in the iOS software supply chain has received limited attention. In this paper, we focus on CocoaPods, one of the most widely used dependency management systems for iOS app development, but also examine the security of Carthage and Swift Package Manager (SwiftPM). We demonstrate that iOS apps expose internal package names and versions. Attackers can exploit this leakage to register previously unclaimed dependencies in CocoaPods, enabling remote code execution (RCE) on developer machines and build servers. Additionally, we show that attackers can compromise dependencies by reclaiming abandoned domains and GitHub URLs. Analyzing a dataset of 9,212 apps, we quantify how many apps are susceptible to these vulnerabilities. Further, we inspect the use of vulnerable dependencies within public GitHub repositories. Our findings reveal that popular apps disclose internal dependency information, enabling dependency confusion attacks. Furthermore, we show that hijacking a single CocoaPod library through an abandoned domain could compromise 63 iOS apps, affecting millions of users. Finally, we compare iOS dependency management systems with Cargo, Go modules, Maven, npm, and pip to discuss mitigation strategies for the identified threats.

A computer is nothing but a device that processes the instructions supplied to it. However, as computers evolved, the instructions or codes started to be more complicated. As computers started to be used by non-technical people, it became imperative that the users be able to use the machine without having underlying knowledge of the code or the hardware. And operating system became the backbone for translating the inputs from the user to actual operation on the hardware. With the increasing complexity and the choices of operating system, it became clear that different groups of people, especially in an enterprise scenario, required different operating systems. Installing them all on a single machine, for shared computers became a difficult task, giving rise to network-based booting. But network-based booting was confined to only wired connectivity, keeping it restricted to very small geographical areas. The proposed system, /dev/SDB, is aimed at creating a standard where any user, anyone on the globe, can access the operating system authorized to them without having to be on the corporate network. It aims to offer the same over Wi-Fi as well as cellular connectivity, ensuring employees can truly work from anywhere, while following the policies for operating systems and without redundant hardware.

This paper critically examines the device identification process using machine learning, addressing common pitfalls in existing literature. We analyze the trade-offs between identification methods (unique vs. class based), data heterogeneity, feature extraction challenges, and evaluation metrics. By highlighting specific errors, such as improper data augmentation and misleading session identifiers, we provide a robust guideline for researchers to enhance the reproducibility and generalizability of IoT security models.

Mobile devices rely on Trusted Execution Environments (TEEs) to execute security-critical code and protect sensitive assets. This security-critical code is modularized in components known as Trusted Applications (TAs). Vulnerabilities in TAs can compromise the TEE and, thus, the entire system. However, the closed-source nature and fragmentation of mobile TEEs severely hinder dynamic analysis of TAs, limiting testing efforts to mostly static analyses. This paper presents TÄMU, a rehosting platform enabling dynamic analysis of TAs, specifically fuzzing and debugging, by interposing their execution at the API layer. To scale to many TAs across different TEEs, TÄMU leverages the standardization of TEE APIs, driven by the GlobalPlatform specifications. For the remaining TEE-specific APIs not shared across different TEEs, TÄMU introduces the notion of greedy high-level emulation, a technique that allows prioritizing manual rehosting efforts based on the potential coverage gain during fuzzing. We implement TÄMU and use it to emulate 67 TAs across four TEEs. Our fuzzing campaigns yielded 17 zero-day vulnerabilities across 11 TAs. These results indicate a deficit of dynamic analysis capabilities across the TEE ecosystem, where not even vendors with source code unlocked these capabilities for themselves. TÄMU promises to close this gap by bringing effective and practical dynamic analysis to the mobile TEE domain.

Private Set Multi-Party Computations are protocols that allow parties to jointly and securely compute functions: apart from what is deducible from the output of the function, the input sets are kept private. Then, a Private Set Union (PSU), resp. Intersection (PSI), is a protocol that allows parties to jointly compute the union, resp. the intersection, between their private sets. Now a structured PSI, is a PSI where some structure of the sets can allow for more efficient protocols. For instance in Fuzzy PSI, elements only need to be close enough, instead of equal, to be part of the intersection. We present in this paper, Fuzzy PSU protocols (FPSU), able to efficiently take into account approximations in the union. For this, we introduce a new efficient sub-protocol, called Oblivious Key Homomorphic Encryption Retrieval (OKHER), improving on Oblivious Key-Value Retrieval (OKVR) techniques in our setting. In the fuzzy context, the receiver set $X=\{x_i\}_{1..n}$ is replaced by ${\mathcal B}_\delta (X)$, the union of $n$ balls of dimension $d$ with radius $\delta$, centered at the $x_i$. The sender set is just its $m$ points of dimension $d$. Then the FPSU functionality corresponds to $X \sqcup \{y \in Y, y \notin {\mathcal B}_\delta (X)\}$. Thus, we formally define the FPSU functionality and security properties, and propose several protocols tuned to the patterns of the balls using the $l_\infty$ distance. Using our OKHER routine and homomorphic encryption, we are for instance able to obtain a FPSU protocols with an asymptotic communication volume bound ranging from $O(dm\log(\delta{n}))$ to $O(d^2m\log(\delta^2n))$, depending on the receiver data set structure.

As Open Radio Access Network (O-RAN) deployments expand and adversaries adopt 'store-now, decrypt-later' strategies, operators need empirical data on the cost of migrating critical control interfaces to post-quantum cryptography (PQC). This paper experimentally evaluates the impact of integrating a NIST-aligned module-lattice KEM (ML-KEM, CRYSTALS-Kyber) into IKEv2/IPsec protecting the E2 interface between the 5G Node B (gNB) and the Near-Real-Time RAN Intelligent Controller (Near-RT RIC). Using an open-source testbed built from srsRAN, Open5GS, FlexRIC and strongSwan (with liboqs), we compare three configurations: no IPsec, classical ECDH-based IPsec, and ML-KEM-based IPsec. The study focuses on IPsec tunnel-setup latency and the runtime behaviour of Near-RT RIC xApps under realistic signalling workloads. Results from repeated, automated runs show that ML-KEM integration adds a small overhead to tunnel establishment, which is approximately 3~5 ms in comparison to classical IPsec, while xApp operation and RIC control loops remain stable in our experiments. These findings indicate that ML-KEM based IPsec on the E2 interface is practically feasible and inform quantum-safe migration strategies for O-RAN deployments.

The current digital era, driven by growing threats to data security, requires a robust image encryption technique. Classical encryption algorithms suffer from a trade-off among security, image fidelity, and computational efficiency. This paper aims to enhance the performance and efficiency of image encryption. This is done by proposing Fractal encryption based on Fourier transforms as a new method of image encryption, leveraging state-of-the-art technology. The new approach considered here intends to enhance both security and efficiency in image encryption by comparing Fractal Encryption with basic methods. The suggested system also aims to optimise encryption/ decryption times and preserve image quality. This paper provides an introduction to Image Encryption using the fractal-based method, its mathematical formulation, and its comparative efficiency against publicly known traditional encryption methods. As a result, after filling the gaps identified in previous research, it has significantly improved both its encryption/decryption time and image fidelity compared to other techniques. In this paper, directions for future research and possible improvements are outlined for attention.

Recently, a novel peer sampling protocol, Elevator, was introduced to construct network topologies tailored for emerging decentralized applications such as federated learning and blockchain. Elevator builds hub-based topologies in a fully decentralized manner, randomly selecting hubs among participating nodes. These hubs, acting as central nodes connected to the entire network, can be leveraged to accelerate message dissemination. Simulation results have shown that Elevator converges rapidly (within 3–4 cycles) and exhibits robustness against crash failures and churn. However, its resilience to Byzantine adversaries has not been investigated. In this work, we provide the first evaluation of Elevator under Byzantine adversaries and show that even a small fraction (2%) of Byzantine nodes is sufficient to subvert the network. As a result, we introduce LIFT, a new protocol that extends Elevator by employing a cryptographically secure pseudo-random number generator (PRNG) for hub selection, thereby mitigating Byzantine manipulation. In contrast, LIFT withstands adversarial infiltration and remains robust with up to 10% Byzantine nodes. These results highlight the necessity of secure randomness in decentralized hub formation and position LIFT as a more reliable building block for Byzantine-resilient decentralized systems.

Ransomware has become one of the most serious cybersecurity threats causing major financial losses and operational disruptions worldwide.Traditional detection methods such as static analysis, heuristic scanning and behavioral analysis often fall short when used alone. To address these limitations, this paper presents multimodal multi agent ransomware analysis framework designed for ransomware classification. Proposed multimodal multiagent architecture combines information from static, dynamic and network sources. Each data type is handled by specialized agent that uses auto encoder based feature extraction. These representations are then integrated through a fusion agent. After that fused representation are used by transformer based classifier. It identifies the specific ransomware family. The agents interact through an interagent feedback mechanism that iteratively refines feature representations by suppressing low confidence information. The framework was evaluated on large scale datasets containing thousands of ransomware and benign samples. Multiple experiments were conducted on ransomware dataset. It outperforms single modality and nonadaptive fusion baseline achieving improvement of up to 0.936 in Macro-F1 for family classification and reducing calibration error. Over 100 epochs, the agentic feedback loop displays a stable monotonic convergence leading to over +0.75 absolute improvement in terms of agent quality and a final composite score of around 0.88 without fine tuning of the language models. Zeroday ransomware detection remains family dependent on polymorphism and modality disruptions. Confidence aware abstention enables reliable real world deployment by favoring conservativeand trustworthy decisions over forced classification. The findings indicate that proposed approach provides a practical andeffective path toward improving real world ransomware defense systems.

Machine unlearning is an emerging technique that aims to remove the influence of specific data from trained models, thereby enhancing privacy protection. However, recent research has uncovered critical privacy vulnerabilities, showing that adversaries can exploit unlearning inversion to reconstruct data that was intended to be erased. Despite the severity of this threat, dedicated defenses remain lacking. To address this gap, we propose UnlearnShield, the first defense specifically tailored to counter unlearning inversion. UnlearnShield introduces directional perturbations in the cosine representation space and regulates them through a constraint module to jointly preserve model accuracy and forgetting efficacy, thereby reducing inversion risk while maintaining utility. Experiments demonstrate that it achieves a good trade-off among privacy protection, accuracy, and forgetting.

Latent-based watermarks, integrated into the generation process of latent diffusion models (LDMs), simplify detection and attribution of generated images. However, recent black-box forgery attacks, where an attacker needs at least one watermarked image and black-box access to the provider's model, can embed the provider's watermark into images not produced by the provider, posing outsized risk to provenance and trust. We propose SemBind, the first defense framework for latent-based watermarks that resists black-box forgery by binding latent signals to image semantics via a learned semantic masker. Trained with contrastive learning, the masker yields near-invariant codes for the same prompt and near-orthogonal codes across prompts; these codes are reshaped and permuted to modulate the target latent before any standard latent-based watermark. SemBind is generally compatible with existing latent-based watermarking schemes and keeps image quality essentially unchanged, while a simple mask-ratio parameter offers a tunable trade-off between anti-forgery strength and robustness. Across four mainstream latent-based watermark methods, our SemBind-enabled anti-forgery variants markedly reduce false acceptance under black-box forgery while providing a controllable robustness-security balance.

Phishing continues to be one of the most prevalent attack vectors, making accurate classification of phishing URLs essential. Recently, large language models (LLMs) have demonstrated promising results in phishing URL detection. However, their reasoning capabilities that enabled such performance remain underexplored. To this end, in this paper, we propose a Least-to-Most prompting framework for phishing URL detection. In particular, we introduce an "answer sensitivity" mechanism that guides Least-to-Most's iterative approach to enhance reasoning and yield higher prediction accuracy. We evaluate our framework using three URL datasets and four state-of-the-art LLMs, comparing against a one-shot approach and a supervised model. We demonstrate that our framework outperforms the one-shot baseline while achieving performance comparable to that of the supervised model, despite requiring significantly less training data. Furthermore, our in-depth analysis highlights how the iterative reasoning enabled by Least-to-Most, and reinforced by our answer sensitivity mechanism, drives these performance gains. Overall, we show that this simple yet powerful prompting strategy consistently outperforms both one-shot and supervised approaches, despite requiring minimal training or few-shot guidance. Our experimental setup can be found in our Github repository github.sydney.edu.au/htri0928/least-to-most-phishing-detection.

The increasing integration of AI agents into cyber-physical systems (CPS) introduces new security risks that extend beyond traditional cyber or physical threat models. Recent advances in generative AI enable deepfake and semantic manipulation attacks that can compromise agent perception, reasoning, and interaction with the physical environment, while emerging protocols such as the Model Context Protocol (MCP) further expand the attack surface through dynamic tool use and cross-domain context sharing. This survey provides a comprehensive review of security threats targeting AI agents in CPS, with a particular focus on environmental interactions, deepfake-driven attacks, and MCP-mediated vulnerabilities. We organize the literature using the SENTINEL framework, a lifecycle-aware methodology that integrates threat characterization, feasibility analysis under CPS constraints, defense selection, and continuous validation. Through an end-to-end case study grounded in a real-world smart grid deployment, we quantitatively illustrate how timing, noise, and false-positive costs constrain deployable defenses, and why detection mechanisms alone are insufficient as decision authorities in safety-critical CPS. The survey highlights the role of provenance- and physics-grounded trust mechanisms and defense-in-depth architectures, and outlines open challenges toward trustworthy AI-enabled CPS.

Always-on hardware Trojans (HTs) pose a critical risk to trusted microelectronics, yet most side-channel detection methods rely on unavailable golden references. We present a reference-free approach that combines time-frequency EM analysis with Gaussian Mixture Models (GMMs). By applying Short-Time Fourier Transform (STFT) at multiple window sizes, we show that HT-free circuits exhibit fluctuating statistical structure, while always-on HTs leave persistent footprints with fewer, more consistent mixture components. Results on AES-128 demonstrate feasibility without requiring reference models.

With the widespread adoption of Large Language Models (LLMs) and increasingly stringent privacy regulations, protecting data privacy in LLMs has become essential, especially for privacy-sensitive applications. Membership Inference Attacks (MIAs) attempt to determine whether a specific data sample was included in the model training/fine-tuning dataset, posing serious privacy risks. However, most existing MIA techniques against LLMs rely on sequence-level aggregated prediction statistics, which fail to distinguish prediction improvements caused by generalization from those caused by memorization, leading to low attack effectiveness. To address this limitation, we propose a novel membership inference approach that captures the token-level probabilities for low-confidence (hard) tokens, where membership signals are more pronounced. By comparing token-level probability improvements at hard tokens between a fine-tuned target model and a pre-trained reference model, HT-MIA isolates strong and robust membership signals that are obscured by prior MIA approaches. Extensive experiments on both domain-specific medical datasets and general-purpose benchmarks demonstrate that HT-MIA consistently outperforms seven state-of-the-art MIA baselines. We further investigate differentially private training as an effective defense mechanism against MIAs in LLMs. Overall, our HT-MIA framework establishes hard-token based analysis as a state-of-the-art foundation for advancing membership inference attacks and defenses for LLMs.

Morphing techniques generate artificial biometric samples that combine features from multiple individuals, allowing each contributor to be verified against a single enrolled template. While extensively studied in face recognition, this vulnerability remains largely unexplored in voice biometrics. Prior work on voice morphing is computationally expensive, non-scalable, and limited to acoustically similar identity pairs, constraining practical deployment. Moreover, existing sound-morphing methods target audio textures, music, or environmental sounds and are not transferable to voice identity manipulation. We propose VoxMorph, a zero-shot framework that produces high-fidelity voice morphs from as little as five seconds of audio per subject without model retraining. Our method disentangles vocal traits into prosody and timbre embeddings, enabling fine-grained interpolation of speaking style and identity. These embeddings are fused via Spherical Linear Interpolation (Slerp) and synthesized using an autoregressive language model coupled with a Conditional Flow Matching network. VoxMorph achieves state-of-the-art performance, delivering a 2.6x gain in audio quality, a 73% reduction in intelligibility errors, and a 67.8% morphing attack success rate on automated speaker verification systems under strict security thresholds. This work establishes a practical and scalable paradigm for voice morphing with significant implications for biometric security. The code and dataset are available on our project page: https://vcbsl.github.io/VoxMorph/

Even though demonstrating extraordinary capabilities in code generation and software issue resolving, AI agents' capabilities in the full software DevOps cycle are still unknown. Different from pure code generation, handling the DevOps cycle in real-world software, including developing, deploying, and managing, requires analyzing large-scale projects, understanding dynamic program behaviors, leveraging domain-specific tools, and making sequential decisions. However, existing benchmarks focus on isolated problems and lack environments and tool interfaces for DevOps. We introduce DevOps-Gym, the first end-to-end benchmark for evaluating AI agents across core DevOps workflows: build and configuration, monitoring, issue resolving, and test generation. DevOps-Gym includes 700+ real-world tasks collected from 30+ projects in Java and Go. We develop a semi-automated data collection mechanism with rigorous and non-trivial expert efforts in ensuring the task coverage and quality. Our evaluation of state-of-the-art models and agents reveals fundamental limitations: they struggle with issue resolving and test generation in Java and Go, and remain unable to handle new tasks such as monitoring and build and configuration. These results highlight the need for essential research in automating the full DevOps cycle with AI agents.

We study the problem of covert quickest change detection in a discrete-time setting, where a sequence of observations undergoes a distributional change at an unknown time. Unlike classical formulations, we consider a covert adversary who has knowledge of the detector's false alarm constraint parameter $\gamma$ and selects a stationary post-change distribution that depends on it, seeking to remain undetected for as long as possible. Building on the theoretical foundations of the CuSum procedure, we rigorously characterize the asymptotic behavior of the average detection delay (ADD) and the average time to false alarm (AT2FA) when the post-change distribution converges to the pre-change distribution as $\gamma \to \infty$. Our analysis establishes exact asymptotic expressions for these quantities, extending and refining classical results that no longer hold in this regime. We identify the critical scaling laws governing covert behavior and derive explicit conditions under which an adversary can maintain covertness, defined by ADD = $\Theta(\gamma)$, whereas in the classical setting, ADD grows only as $\mathcal{O}(\log \gamma)$. In particular, for Gaussian and Exponential models under adversarial perturbations of their respective parameters, we asymptotically characterize ADD as a function of the Kullback–Leibler divergence between the pre- and post-change distributions and $\gamma$.

As large language models (LLMs) move from research prototypes to enterprise systems, their security vulnerabilities pose serious risks to data privacy and system integrity. This study benchmarks various Llama model variants against the OWASP Top 10 for LLM Applications framework, evaluating threat detection accuracy, response safety, and computational overhead. Using the FABRIC testbed with NVIDIA A30 GPUs, we tested five standard Llama models and five Llama Guard variants on 100 adversarial prompts covering ten vulnerability categories. Our results reveal significant differences in security performance: the compact Llama-Guard-3-1B model achieved the highest detection rate of 76% with minimal latency (0.165s per test), whereas base models such as Llama-3.1-8B failed to detect threats (0% accuracy) despite longer inference times (0.754s). We observe an inverse relationship between model size and security effectiveness, suggesting that smaller, specialized models often outperform larger general-purpose ones in security tasks. Additionally, we provide an open-source benchmark dataset including adversarial prompts, threat labels, and attack metadata to support reproducible research in AI security, [1].

What is the AGI in Offensive Security? One can break it down into two questions : (1) any offensive security tasks could be reduced into symbolic language manipulation (language representation + reasoning), (2) powerful language model (LLM) are enough to "deal with" any symbolic language manipulation. This paper can formally model a target system as a state machine and a hacker as an interactive symbolic agent. And it shows that every interaction in an offensive engagement can be encoded as a finite string. This paper provides definitions, short lemmas, and open discussion.

European digital identity initiatives are grounded in regulatory frameworks designed to ensure interoperability and robust, harmonized security standards. The evolution of these frameworks culminates in eIDAS 2.0, whose origins trace back to the Electronic Signatures Directive 1999/93/EC, the first EU-wide legal foundation for the use of electronic signatures in cross-border electronic transactions. As technological capabilities advanced, the initial eIDAS 1.0 framework was increasingly criticized for its limitations and lack of comprehensiveness. Emerging decentralized approaches further exposed these shortcomings and introduced the possibility of integrating innovative identity paradigms, such as Self-Sovereign Identity (SSI) models. In this article, we contribute to the ongoing legal and policy debate on the European Digital Identity Framework by analyzing key provisions of eIDAS 2.0 and its accompanying recitals, drawing on a systematic literature review guided by defined Research Questions (RQ). This work employs a structured methodological approach that combines descriptive and comparative analysis, systematic gap analysis supported by a defined scoring matrix, and normative analysis to evaluate the compatibility of SSI properties with eIDAS 2.0 regulation, as operationalized via its Architecture and Reference Framework (ARF). Furthermore, we assess the ARF's guidelines and examine the extent to which it aligns with SSI. The analysis adopts a complementary perspective demonstrating how the regulation can be further developed to better support SSI in the future by identifying existing limitations and potential adoption opportunities within the current legal foundations of the framework.

Large language models (LLMs) are increasingly paired with activation-based monitoring to detect and prevent harmful behaviors that may not be apparent at the surface-text level. However, existing activation safety approaches, trained on broad misuse datasets, struggle with poor precision, limited flexibility, and lack of interpretability. This paper introduces a new paradigm: rule-based activation safety, inspired by rule-sharing practices in cybersecurity. We propose modeling activations as cognitive elements (CEs), fine-grained, interpretable factors such as 'making a threat' and 'payment processing', that can be composed to capture nuanced, domain-specific behaviors with higher precision. Building on this representation, we present a practical framework that defines predicate rules over CEs and detects violations in real time. This enables practitioners to configure and update safeguards without retraining models or detectors, while supporting transparency and auditability. Our results show that compositional rule-based activation safety improves precision, supports domain customization, and lays the groundwork for scalable, interpretable, and auditable AI governance. We open source GAVEL and introduce GAVEL Studio, an interactive rule authoring and management tool. Code and datasets are available at github.com/Offensive-AI-Lab/gavel.

The dual offensive and defensive utility of Large Language Models (LLMs) highlights a critical gap in AI security: the lack of unified frameworks for dynamic, iterative adversarial adaptation hardening. To bridge this gap, we propose the Red Team vs. Blue Team (RvB) framework, formulated as a training-free, sequential, imperfect-information game. In this process, the Red Team exposes vulnerabilities, driving the Blue Team to learning effective solutions without parameter updates. We validate our framework across two challenging domains: dynamic code hardening against CVEs and guardrail optimization against jailbreaks. Our empirical results show that this interaction compels the Blue Team to learn fundamental defensive principles, leading to robust remediations that are not merely overfitted to specific exploits. RvB achieves Defense Success Rates of 90% and 45% across the respective tasks while maintaining near 0% False Positive Rates, significantly surpassing baselines. This work establishes the iterative adversarial interaction framework as a practical paradigm that automates the continuous hardening of AI systems.

User authentication and fraud detection face growing challenges as digital systems expand and adversaries adopt increasingly sophisticated tactics. Traditional knowledge-based authentication remains rigid, requiring exact word-for-word string matches that fail to accommodate natural human memory and linguistic variation. Meanwhile, fraud-detection pipelines struggle to keep pace with rapidly evolving scam behaviors, leading to high false-positive rates and frequent retraining cycles required. This work introduces two complementary LLM-enabled solutions, namely, an LLM-assisted authentication mechanism that evaluates semantic correctness rather than exact wording, supported by document segmentation and a hybrid scoring method combining LLM judgement with cosine-similarity metrics and a RAG-based fraud-detection pipeline that grounds LLM reasoning in curated evidence to reduce hallucinations and adapt to emerging scam patterns without model retraining. Experiments show that the authentication system accepts 99.5% of legitimate non-exact answers while maintaining a 0.1% false-acceptance rate, and that the RAG-enhanced fraud detection reduces false positives from 17.2% to 3.5%. Together, these findings demonstrate that LLMs can significantly improve both usability and robustness in security workflows, offering a more adaptive , explainable, and human-aligned approach to authentication and fraud detection.

We study the feasibility, profitability, and prevalence of sandwich attacks on Ethereum rollups with private mempools. First, we extend a formal model of optimal front- and back-run sizing, relating attack profitability to victim trade volume, liquidity depth, and slippage bounds. We complement it with an execution-feasibility model that quantifies co-inclusion constraints under private mempools. Second, we examine execution constraints in the absence of builder markets: without guaranteed atomic inclusion, attackers must rely on sequencer ordering, redundant submissions, and priority fee placement, which renders sandwiching probabilistic rather than deterministic. Third, using transaction-level data from major rollups, we show that naive heuristics overstate sandwich activity. We find that the majority of flagged patterns are false positives and that the median net return for these attacks is negative. Our results suggest that sandwiching, while endemic and profitable on Ethereum L1, is rare, unprofitable, and largely absent in rollups with private mempools. These findings challenge prevailing assumptions, refine measurement of MEV in L2s, and inform the design of sequencing policies.

Always-on sensing of AI applications on AR glasses makes traditional permission techniques ill-suited for context-dependent visual data, especially within home environments. The home presents a highly challenging privacy context due to the high density of sensitive objects, and the frequent presence of non-consenting family members, and the intimate nature of daily routines, making it a critical focus area for scalable privacy control mechanisms. Existing fine-grained controls, while offering nuanced choices, are inefficient for managing multiple private objects. We propose VisGuardian, a fine-grained content-based visual permission technique for AR glasses. VisGuardian features a group-based control mechanism that enables users to efficiently manage permissions for multiple private objects. VisGuardian detects objects using YOLO and adopts a pre-classified schema to group them. By selecting a single object, users can efficiently obscure groups of related objects based on criteria including privacy sensitivity, object category, or spatial proximity. A technical evaluation shows VisGuardian achieves mAP50 of 0.6704 with only 14.0 ms latency and a 1.7% increase in battery consumption per hour. Furthermore, a user study (N=24) comparing VisGuardian to slider-based and object-based baselines found it to be significantly faster for setting permissions and was preferred by users for its efficiency, effectiveness, and ease of use.