Traditional white-box methods for creating adversarial perturbations against LLMs typically rely only on gradient computation from the targeted model, ignoring the internal mechanisms responsible for attack success or failure. Conversely, interpretability studies that analyze these internal mechanisms lack practical applications beyond runtime interventions. We bridge this gap by introducing a novel white-box approach that leverages mechanistic interpretability techniques to craft practical adversarial inputs. Specifically, we first identify acceptance subspaces - sets of feature vectors that do not trigger the model's refusal mechanisms - then use gradient-based optimization to reroute embeddings from refusal subspaces to acceptance subspaces, effectively achieving jailbreaks. This targeted approach significantly reduces computation cost, achieving attack success rates of 80-95\% on state-of-the-art models including Gemma2, Llama3.2, and Qwen2.5 within minutes or even seconds, compared to existing techniques that often fail or require hours of computation. We believe this approach opens a new direction for both attack research and defense development. Furthermore, it showcases a practical application of mechanistic interpretability where other methods are less efficient, which highlights its utility. The code and generated datasets are available at https://github.com/Sckathach/subspace-rerouting.

As national security institutions increasingly integrate Artificial Intelligence (AI) into decision-making and content generation processes, understanding the inherent biases of large language models (LLMs) is crucial. This study presents a novel benchmark designed to evaluate the biases and preferences of seven prominent foundation models-Llama 3.1 8B Instruct, Llama 3.1 70B Instruct, GPT-4o, Gemini 1.5 Pro-002, Mixtral 8x22B, Claude 3.5 Sonnet, and Qwen2 72B-in the context of international relations (IR). We designed a bias discovery study around core topics in IR using 400-expert crafted scenarios to analyze results from our selected models. These scenarios focused on four topical domains including: military escalation, military and humanitarian intervention, cooperative behavior in the international system, and alliance dynamics. Our analysis reveals noteworthy variation among model recommendations based on scenarios designed for the four tested domains. Particularly, Qwen2 72B, Gemini 1.5 Pro-002 and Llama 3.1 8B Instruct models offered significantly more escalatory recommendations than Claude 3.5 Sonnet and GPT-4o models. All models exhibit some degree of country-specific biases, often recommending less escalatory and interventionist actions for China and Russia compared to the United States and the United Kingdom. These findings highlight the necessity for controlled deployment of LLMs in high-stakes environments, emphasizing the need for domain-specific evaluations and model fine-tuning to align with institutional objectives.

With LLM usage rapidly increasing, their vulnerability to jailbreaks that create harmful outputs are a major security risk. As new jailbreaking strategies emerge and models are changed by fine-tuning, continuous testing for security vulnerabilities is necessary. Existing Red Teaming methods fall short in cost efficiency, attack success rate, attack diversity, or extensibility as new attack types emerge. We address these challenges with Modular And Diverse Malicious Attack MiXtures (MAD-MAX) for Automated LLM Red Teaming. MAD-MAX uses automatic assignment of attack strategies into relevant attack clusters, chooses the most relevant clusters for a malicious goal, and then combines strategies from the selected clusters to achieve diverse novel attacks with high attack success rates. MAD-MAX further merges promising attacks together at each iteration of Red Teaming to boost performance and introduces a similarity filter to prune out similar attacks for increased cost efficiency. The MAD-MAX approach is designed to be easily extensible with newly discovered attack strategies and outperforms the prominent Red Teaming method Tree of Attacks with Pruning (TAP) significantly in terms of Attack Success Rate (ASR) and queries needed to achieve jailbreaks. MAD-MAX jailbreaks 97% of malicious goals in our benchmarks on GPT-4o and Gemini-Pro compared to TAP with 66%. MAD-MAX does so with only 10.9 average queries to the target LLM compared to TAP with 23.3. WARNING: This paper contains contents which are offensive in nature.

Machine learning models were shown to be vulnerable to model stealing attacks, which lead to intellectual property infringement. Among other methods, substitute model training is an all-encompassing attack applicable to any machine learning model whose behaviour can be approximated from input-output queries. Whereas prior works mainly focused on improving the performance of substitute models by, e.g. developing a new substitute training method, there have been only limited ablation studies on the impact the attacker's strength has on the substitute model's performance. As a result, different authors came to diverse, sometimes contradicting, conclusions. In this work, we exhaustively examine the ambivalent influence of different factors resulting from varying the attacker's capabilities and knowledge on a substitute training attack. Our findings suggest that some of the factors that have been considered important in the past are, in fact, not that influential; instead, we discover new correlations between attack conditions and success rate. In particular, we demonstrate that better-performing target models enable higher-fidelity attacks and explain the intuition behind this phenomenon. Further, we propose to shift the focus from the complexity of target models toward the complexity of their learning tasks. Therefore, for the substitute model, rather than aiming for a higher architecture complexity, we suggest focusing on getting data of higher complexity and an appropriate architecture. Finally, we demonstrate that even in the most limited data-free scenario, there is no need to overcompensate weak knowledge with millions of queries. Our results often exceed or match the performance of previous attacks that assume a stronger attacker, suggesting that these stronger attacks are likely endangering a model owner's intellectual property to a significantly higher degree than shown until now.

Access to real-world healthcare data is limited by stringent privacy regulations and data imbalances, hindering advancements in research and clinical applications. Synthetic data presents a promising solution, yet existing methods often fail to ensure the realism, utility, and calibration essential for robust survival analysis. Here, we introduce Masked Clinical Modelling (MCM), an attention-based framework capable of generating high-fidelity synthetic datasets that preserve critical clinical insights, such as hazard ratios, while enhancing survival model calibration. Unlike traditional statistical methods like SMOTE and machine learning models such as VAEs, MCM supports both standalone dataset synthesis for reproducibility and conditional simulation for targeted augmentation, addressing diverse research needs. Validated on a chronic kidney disease electronic health records dataset, MCM reduced the general calibration loss over the entire dataset by 15%; and MCM reduced a mean calibration loss by 9% across 10 clinically stratified subgroups, outperforming 15 alternative methods.

Traditional online content moderation systems struggle to classify modern multimodal means of communication, such as memes, a highly nuanced and information-dense medium. This task is especially hard in a culturally diverse society like Singapore, where low-resource languages are used and extensive knowledge on local context is needed to interpret online content. We curate a large collection of 112K memes labeled by GPT-4V for fine-tuning a VLM to classify offensive memes in Singapore context. We show the effectiveness of fine-tuned VLMs on our dataset, and propose a pipeline containing OCR, translation and a 7-billion parameter-class VLM. Our solutions reach 80.62% accuracy and 0.8192 AUROC on a held-out test set, and can greatly aid human in moderating online contents. The dataset, code, and model weights have been open-sourced at https://github.com/aliencaocao/vlm-for-memes-aisg.

Multimodal large language models (MLLMs) equipped with Retrieval Augmented Generation (RAG) leverage both their rich parametric knowledge and the dynamic, external knowledge to excel in tasks such as Question Answering. While RAG enhances MLLMs by grounding responses in query-relevant external knowledge, this reliance poses a critical yet underexplored safety risk: knowledge poisoning attacks, where misinformation or irrelevant knowledge is intentionally injected into external knowledge bases to manipulate model outputs to be incorrect and even harmful. To expose such vulnerabilities in multimodal RAG, we propose MM-PoisonRAG, a novel knowledge poisoning attack framework with two attack strategies: Localized Poisoning Attack (LPA), which injects query-specific misinformation in both text and images for targeted manipulation, and Globalized Poisoning Attack (GPA) to provide false guidance during MLLM generation to elicit nonsensical responses across all queries. We evaluate our attacks across multiple tasks, models, and access settings, demonstrating that LPA successfully manipulates the MLLM to generate attacker-controlled answers, with a success rate of up to 56% on MultiModalQA. Moreover, GPA completely disrupts model generation to 0% accuracy with just a single irrelevant knowledge injection. Our results highlight the urgent need for robust defenses against knowledge poisoning to safeguard multimodal RAG frameworks.

Despite a plethora of anomaly detection models developed over the years, their ability to generalize to unseen anomalies remains an issue, particularly in critical systems. This paper aims to address this challenge by introducing Swift Hydra, a new framework for training an anomaly detection method based on generative AI and reinforcement learning (RL). Through featuring an RL policy that operates on the latent variables of a generative model, the framework synthesizes novel and diverse anomaly samples that are capable of bypassing a detection model. These generated synthetic samples are, in turn, used to augment the detection model, further improving its ability to handle challenging anomalies. Swift Hydra also incorporates Mamba models structured as a Mixture of Experts (MoE) to enable scalable adaptation of the number of Mamba experts based on data complexity, effectively capturing diverse feature distributions without increasing the model's inference time. Empirical evaluations on ADBench benchmark demonstrate that Swift Hydra outperforms other state-of-the-art anomaly detection models while maintaining a relatively short inference time. From these results, our research highlights a new and auspicious paradigm of integrating RL and generative AI for advancing anomaly detection.

The US Department of Justice wants Google to sell off its Chrome browser as part of its final remedy proposal in a landmark antitrust case. The proposal, filed Friday afternoon, says that Google must "promptly and fully divest Chrome, along with any assets or services necessary to successfully complete the divestiture, to a buyer approved by the Plaintiffs in their sole discretion, subject to terms that the Court and Plaintiffs approve." It also would require Google to stop paying partners for preferential treatment of its search engine. The DOJ also demands that Google provide prior notification of any new joint venture, collaboration, or partnership with any company that competes with Google in search or in search text ads. However, the company no longer has to divest its artificial intelligence investments, which was part of an initial set of recommendations issued by the plaintiffs last November.

Elon Musk's so-called Department of Government Efficiency has deployed a proprietary chatbot called GSAi to 1,500 federal workers at the General Services Administration, WIRED has confirmed. The move to automate tasks previously done by humans comes as DOGE continues its purge of the federal workforce. GSAi is meant to support "general" tasks, similar to commercial tools like ChatGPT or Anthropic's Claude. It is tailored in a way that makes it safe for government use, a GSA worker tells WIRED. The DOGE team hopes to eventually use it to analyze contract and procurement data, WIRED previously reported.