Multi-Agent Debate (MAD), leveraging collaborative interactions among Large Language Models (LLMs), aim to enhance reasoning capabilities in complex tasks. However, the security implications of their iterative dialogues and role-playing characteristics, particularly susceptibility to jailbreak attacks eliciting harmful content, remain critically underexplored. This paper systematically investigates the jailbreak vulnerabilities of four prominent MAD frameworks built upon leading commercial LLMs (GPT-4o, GPT-4, GPT-3.5-turbo, and DeepSeek) without compromising internal agents. We introduce a novel structured prompt-rewriting framework specifically designed to exploit MAD dynamics via narrative encapsulation, role-driven escalation, iterative refinement, and rhetorical obfuscation. Our extensive experiments demonstrate that MAD systems are inherently more vulnerable than single-agent setups. Crucially, our proposed attack methodology significantly amplifies this fragility, increasing average harmfulness from 28.14% to 80.34% and achieving attack success rates as high as 80% in certain scenarios. These findings reveal intrinsic vulnerabilities in MAD architectures and underscore the urgent need for robust, specialized defenses prior to real-world deployment.

The transfer-based black-box adversarial attack setting poses the challenge of crafting an adversarial example (AE) on known surrogate models that remain effective against unseen target models. Due to the practical importance of this task, numerous methods have been proposed to address this challenge. However, most previous methods are heuristically designed and intuitively justified, lacking a theoretical foundation. To bridge this gap, we derive a novel transferability bound that offers provable guarantees for adversarial transferability. Our theoretical analysis has the advantages of \textit{(i)} deepening our understanding of previous methods by building a general attack framework and \textit{(ii)} providing guidance for designing an effective attack algorithm. Our theoretical results demonstrate that optimizing AEs toward flat minima over the surrogate model set, while controlling the surrogate-target model shift measured by the adversarial model discrepancy, yields a comprehensive guarantee for AE transferability. The results further lead to a general transfer-based attack framework, within which we observe that previous methods consider only partial factors contributing to the transferability. Algorithmically, inspired by our theoretical results, we first elaborately construct the surrogate model set in which models exhibit diverse adversarial vulnerabilities with respect to AEs to narrow an instantiated adversarial model discrepancy. Then, a \textit{model-Diversity-compatible Reverse Adversarial Perturbation} (DRAP) is generated to effectively promote the flatness of AEs over diverse surrogate models to improve transferability. Extensive experiments on NIPS2017 and CIFAR-10 datasets against various target models demonstrate the effectiveness of our proposed attack.

--Quantum computing has the potential to improve our ability to solve certain optimization problems that are computationally difficult for classical computers, by offering new algorithmic approaches that may provide speedups under specific conditions. In this work, we introduce QAOA-GPT, a generative framework that leverages Generative Pretrained Transformers (GPT) to directly synthesize quantum circuits for solving quadratic unconstrained binary optimization problems, and demonstrate it on the MaxCut problem on graphs. T o diversify the training circuits and ensure their quality, we have generated a synthetic dataset using the adaptive QAOA approach, a method that incrementally builds and optimizes problem-specific circuits. The experiments conducted on a curated set of graph instances demonstrate that QAOA-GPT, generates high quality quantum circuits for new problem instances unseen in the training as well as successfully parametrizes QAOA. Our results show that using QAOA-GPT to generate quantum circuits will significantly decrease both the computational overhead of classical QAOA and adaptive approaches that often use gradient evaluation to generate the circuit and the classical optimization of the circuit parameters. Our work shows that generative AI could be a promising avenue to generate compact quantum circuits in a scalable way. Quantum computing is rapidly emerging technology with significant potential across various domains, including finance [1], chemical simulations [2], material science [3], combinatorial optimization [4], and machine learning [5], among others. V ariational quantum-classical algorithms represent one of the most promising classes of quantum algorithms in different domains, showing potential for both fault-tolerant quantum computers and near-term noisy intermediate-scale quantum (NISQ) devices. The Quantum Approximate Optimization Algorithm (QAOA) [6] and many of its subsequent versions and customizations [7] belong to this class and demonstrate great potential due to their problem/application flexibility and compatibility with various quantum architectures. The original QAOA framework employs a fixed ansatz structure, which can limit expressibility and hinder performance, particularly on near-term quantum devices where circuit depth is limited. This rigid design may not capture the problem-specific features needed for efficient optimization. Such methods as ADAPT -QAOA [8] address this challenge by iteratively constructing the ansatz in a problem-informed manner. At each step, ADAPT -QAOA selects operators from a predefined pool based on their gradient with respect to the cost function, incorporating only those that contribute most significantly to improving the objective.

This paper provides theoretical and empirical comparisons of three recent hierarchical plan repair algorithms: SHOPFixer, IPyHOPPER, and Rewrite. Our theoretical results show that the three algorithms correspond to three different definitions of the plan repair problem, leading to differences in the algorithms' search spaces, the repair problems they can solve, and the kinds of repairs they can make. Understanding these distinctions is important when choosing a repair method for any given application. Building on the theoretical results, we evaluate the algorithms empirically in a series of benchmark planning problems. Our empirical results provide more detailed insight into the runtime repair performance of these systems and the coverage of the repair problems solved, based on algorithmic properties such as replanning, chronological backtracking, and backjumping over plan trees.

As artificial intelligence (AI) reshapes industries and societies, ensuring its trustworthiness -- through mitigating ethical risks like bias, opacity, and accountability deficits -- remains a global challenge. International Organization for Standardization (ISO) AI standards, such as ISO/IEC 24027 and 24368, aim to foster responsible development by embedding fairness, transparency, and risk management into AI systems. However, their effectiveness varies across diverse regulatory landscapes, from the EU' s risk - based AI Act to China's stability - focused measures and the U.S.'s fragmented state - led initiatives. This paper introduces a novel Comparative Risk - Impact Assessment Framework to evaluate how well ISO standards address ethical risks within these cont exts, proposing enhancements to strengthen their global applicability. By mapping ISO standards to the EU AI Act and surveying regulatory frameworks in ten regions -- including the UK, Canada, India, Japan, Singapore, South Korea, and Brazil -- we establish a ba seline for ethical alignment. The framework, applied to case studies in the EU, US - Colorado, and China, reveals gaps: voluntary ISO standards falter in enforcement (e.g., Colorado) and undervalue region - specific risks like privacy (China). We recommend man datory risk audits, region - specific annexes, and a privacy - focused module to enhance ISO's adaptability. This approach not only synthesizes global trends but also offers a replicable tool for aligning standardization with ethical imperatives, fostering int eroperability and trust in AI worldwide. Policymakers and standards bodies can leverage these insights to evolve AI governance, ensuring it meets diverse societal needs as the technology advances .

Governments are starting to impose requirements on AI models based on how much compute was used to train them. For example, the EU AI Act imposes requirements on providers of general-purpose AI with systemic risk, which includes systems trained using greater than $10^{25}$ floating point operations (FLOP). In the United States' AI Diffusion Framework, a training compute threshold of $10^{26}$ FLOP is used to identify "controlled models" which face a number of requirements. We explore how many models such training compute thresholds will capture over time. We estimate that by the end of 2028, there will be between 103-306 foundation models exceeding the $10^{25}$ FLOP threshold put forward in the EU AI Act (90% CI), and 45-148 models exceeding the $10^{26}$ FLOP threshold that defines controlled models in the AI Diffusion Framework (90% CI). We also find that the number of models exceeding these absolute compute thresholds each year will increase superlinearly -- that is, each successive year will see more new models captured within the threshold than the year before. Thresholds that are defined with respect to the largest training run to date (for example, such that all models within one order of magnitude of the largest training run to date are captured by the threshold) see a more stable trend, with a median forecast of 14-16 models being captured by this definition annually from 2025-2028.

Natural Language Processing (NLP) and computational linguistic techniques are increasingly being applied across various domains, yet their use in legal and regulatory tasks remains limited. To address this gap, we develop an efficient bilingual question-answering framework for regulatory documents, specifically the Bangladesh Police Gazettes, which contain both English and Bangla text. Our approach employs modern Retrieval Augmented Generation (RAG) pipelines to enhance information retrieval and response generation. In addition to conventional RAG pipelines, we propose an advanced RAG-based approach that improves retrieval performance, leading to more precise answers. This system enables efficient searching for specific government legal notices, making legal information more accessible. We evaluate both our proposed and conventional RAG systems on a diverse test set on Bangladesh Police Gazettes, demonstrating that our approach consistently outperforms existing methods across all evaluation metrics.

As cyber threats continue to evolve, securing edge networks has become increasingly challenging due to their distributed nature and resource limitations. Many AI-driven threat detection systems rely on complex deep learning models, which, despite their high accuracy, suffer from two major drawbacks: lack of interpretability and high computational cost. Black-box AI models make it difficult for security analysts to understand the reasoning behind their predictions, limiting their practical deployment. Moreover, conventional deep learning techniques demand significant computational resources, rendering them unsuitable for edge devices with limited processing power. To address these issues, this study introduces an Explainable and Lightweight AI (ELAI) framework designed for real-time cyber threat detection in edge networks. Our approach integrates interpretable machine learning algorithms with optimized lightweight deep learning techniques, ensuring both transparency and computational efficiency. The proposed system leverages decision trees, attention-based deep learning, and federated learning to enhance detection accuracy while maintaining explainability. We evaluate ELAI using benchmark cybersecurity datasets, such as CICIDS and UNSW-NB15, assessing its performance across diverse cyberattack scenarios. Experimental results demonstrate that the proposed framework achieves high detection rates with minimal false positives, all while significantly reducing computational demands compared to traditional deep learning methods. The key contributions of this work include: (1) a novel interpretable AI-based cybersecurity model tailored for edge computing environments, (2) an optimized lightweight deep learning approach for real-time cyber threat detection, and (3) a comprehensive analysis of explainability techniques in AI-driven cybersecurity applications.

The exposure of large language models (LLMs) to copyrighted material during pre-training raises concerns about unintentional copyright infringement post deployment. This has driven the development of "copyright takedown" methods, post-training approaches aimed at preventing models from generating content substantially similar to copyrighted ones. While current mitigation approaches are somewhat effective for average-case risks, we demonstrate that they overlook worst-case copyright risks exhibits by the existence of long, verbatim quotes from copyrighted sources. We propose BloomScrub, a remarkably simple yet highly effective inference-time approach that provides certified copyright takedown. Our method repeatedly interleaves quote detection with rewriting techniques to transform potentially infringing segments. By leveraging efficient data sketches (Bloom filters), our approach enables scalable copyright screening even for large-scale real-world corpora. When quotes beyond a length threshold cannot be removed, the system can abstain from responding, offering certified risk reduction. Experimental results show that BloomScrub reduces infringement risk, preserves utility, and accommodates different levels of enforcement stringency with adaptive abstention. Our results suggest that lightweight, inference-time methods can be surprisingly effective for copyright prevention.

Evaluating Large Language Models (LLMs) for safety and security remains a complex task, often requiring users to navigate a fragmented landscape of ad hoc benchmarks, datasets, metrics, and reporting formats. To address this challenge, we present aiXamine, a comprehensive black-box evaluation platform for LLM safety and security. aiXamine integrates over 40 tests (i.e., benchmarks) organized into eight key services targeting specific dimensions of safety and security: adversarial robustness, code security, fairness and bias, hallucination, model and data privacy, out-of-distribution (OOD) robustness, over-refusal, and safety alignment. The platform aggregates the evaluation results into a single detailed report per model, providing a detailed breakdown of model performance, test examples, and rich visualizations. We used aiXamine to assess over 50 publicly available and proprietary LLMs, conducting over 2K examinations. Our findings reveal notable vulnerabilities in leading models, including susceptibility to adversarial attacks in OpenAI's GPT-4o, biased outputs in xAI's Grok-3, and privacy weaknesses in Google's Gemini 2.0. Additionally, we observe that open-source models can match or exceed proprietary models in specific services such as safety alignment, fairness and bias, and OOD robustness. Finally, we identify trade-offs between distillation strategies, model size, training methods, and architectural choices.