Concept erasure in text-to-image diffusion models seeks to remove undesired concepts while preserving overall generative capability. Localized erasure methods aim to restrict edits to the spatial region occupied by the target concept. However, we observe that suppressing a concept can unintentionally weaken semantically related neighbor concepts, reducing fidelity in fine-grained domains. We propose Neighbor-Aware Localized Concept Erasure (NLCE), a training-free framework designed to better preserve neighboring concepts while removing target concepts. It operates in three stages: (1) a spectrally-weighted embedding modulation that attenuates target concept directions while stabilizing neighbor concept representations, (2) an attention-guided spatial gate that identifies regions exhibiting residual concept activation, and (3) a spatially-gated hard erasure that eliminates remaining traces only where necessary. This neighbor-aware pipeline enables localized concept removal while maintaining the surrounding concept neighborhood structure. Experiments on fine-grained datasets (Oxford Flowers, Stanford Dogs) show that our method effectively removes target concepts while better preserving closely related categories. Additional results on celebrity identity, explicit content and artistic style demonstrate robustness and generalization to broader erasure scenarios.

AI agents that interact with the real world through tool calls pose fundamental safety challenges: agents might leak private information, cause unintended side effects, or be manipulated through prompt injection. To address these challenges, we propose to put the agent in a programming-language-based "safety harness": instead of calling tools directly, agents express their intentions as code in a capability-safe language: Scala 3 with capture checking. Capabilities are program variables that regulate access to effects and resources of interest. Scala's type system tracks capabilities statically, providing fine-grained control over what an agent can do. In particular, it enables local purity, the ability to enforce that sub-computations are side-effect-free, preventing information leakage when agents process classified data. We demonstrate that extensible agent safety harnesses can be built by leveraging a strong type system with tracked capabilities. Our experiments show that agents can generate capability-safe code with no significant loss in task performance, while the type system reliably prevents unsafe behaviors such as information leakage and malicious side effects.

Discrete image tokenizers encode visual inputs as sequences of tokens from a finite vocabulary and are gaining popularity in multimodal systems, including encoder-only, encoder-decoder, and decoder-only models. However, unlike CLIP encoders, their vulnerability to adversarial attacks has not been explored. Ours being the first work studying this topic, we first formulate attacks that aim to perturb the features extracted by discrete tokenizers, and thus change the extracted tokens. These attacks are computationally efficient, application-agnostic, and effective across classification, multimodal retrieval, and captioning tasks. Second, to defend against this vulnerability, inspired by recent work on robust CLIP encoders, we fine-tune popular tokenizers with unsupervised adversarial training, keeping all other components frozen. While unsupervised and task-agnostic, our approach significantly improves robustness to both unsupervised and end-to-end supervised attacks and generalizes well to unseen tasks and data. Unlike supervised adversarial training, our approach can leverage unlabeled images, making it more versatile. Overall, our work highlights the critical role of tokenizer robustness in downstream tasks and presents an important step in the development of safe multimodal foundation models.

LLM-based agents execute real-world workflows via tools and memory. These affordances enable ill-intended adversaries to also use these agents to carry out complex misuse scenarios. Existing agent misuse benchmarks largely test single-prompt instructions, leaving a gap in measuring how agents end up helping with harmful or illegal tasks over multiple turns. We introduce STING (Sequential Testing of Illicit N-step Goal execution), an automated red-teaming framework that constructs a step-by-step illicit plan grounded in a benign persona and iteratively probes a target agent with adaptive follow-ups, using judge agents to track phase completion. We further introduce an analysis framework that models multi-turn red-teaming as a time-to-first-jailbreak random variable, enabling analysis tools like discovery curves, hazard-ratio attribution by attack language, and a new metric: Restricted Mean Jailbreak Discovery. Across AgentHarm scenarios, STING yields substantially higher illicit-task completion than single-turn prompting and chat-oriented multi-turn baselines adapted to tool-using agents. In multilingual evaluations across six non-English settings, we find that attack success and illicit-task completion do not consistently increase in lower-resource languages, diverging from common chatbot findings. Overall, STING provides a practical way to evaluate and stress-test agent misuse in realistic deployment settings, where interactions are inherently multi-turn and often multilingual.

Indirect prompt injection attacks threaten AI agents that execute consequential actions, motivating deterministic system-level defenses. Such defenses can provably block unsafe actions by enforcing confidentiality and integrity policies, but currently appear costly: they reduce task completion rates and increase token usage compared to probabilistic defenses. We argue that existing evaluations miss a key benefit of system-level defenses: reduced reliance on human oversight. We introduce autonomy metrics to quantify this benefit: the fraction of consequential actions an agent can execute without human-in-the-loop (HITL) approval while preserving security. To increase autonomy, we design a security-aware agent that (i) introduces richer HITL interactions, and (ii) explicitly plans for both task progress and policy compliance. We implement this agent design atop an existing information-flow control defense against prompt injection and evaluate it on the AgentDojo and WASP benchmarks. Experiments show that this approach yields higher autonomy without sacrificing utility.

Aggregation rules are the cornerstone of distributed (or federated) learning in the presence of adversaries, under the so-called Byzantine threat model. They are also interesting mathematical objects from the point of view of robust mean estimation. The Krum aggregation rule has been extensively studied, and endowed with formal robustness and convergence guarantees. Yet, MultiKrum, a natural extension of Krum, is often preferred in practice for its superior empirical performance, even though no theoretical guarantees were available until now. In this work, we provide the first proof that MultiKrum is a robust aggregation rule, and bound its robustness coefficient. To do so, we introduce $κ^\star$, the optimal *robustness coefficient* of an aggregation rule, which quantifies the accuracy of mean estimation in the presence of adversaries in a tighter manner compared with previously adopted notions of robustness. We then construct an upper and a lower bound on MultiKrum's robustness coefficient. As a by-product, we also improve on the best-known bounds on Krum's robustness coefficient. We show that MultiKrum's bounds are never worse than Krum's, and better in realistic regimes. We illustrate this analysis by an experimental investigation on the quality of the lower bound.

Membership Inference Attacks (MIAs) are widely used to quantify training data memorization and assess privacy risks. Standard evaluation requires repeated retraining, which is computationally costly for large models. One-run methods (single training with randomized data inclusion) and zero-run methods (post hoc evaluation) are often used instead, though their statistical validity remains unclear. To address this gap, we frame MIA evaluation as a causal inference problem, defining memorization as the causal effect of including a data point in the training set. This novel formulation reveals and formalizes key sources of bias in existing protocols: one-run methods suffer from interference between jointly included points, while zero-run evaluations popular for LLMs are confounded by non-random membership assignment. We derive causal analogues of standard MIA metrics and propose practical estimators for multi-run, one-run, and zero-run regimes with non-asymptotic consistency guarantees. Experiments on real-world data show that our approach enables reliable memorization measurement even when retraining is impractical and under distribution shift, providing a principled foundation for privacy evaluation in modern AI systems.

We evaluate the effectiveness of child filtering to prevent the misuse of text-to-image (T2I) models to create child sexual abuse material (CSAM). First, we capture the complexity of preventing CSAM generation using a game-based security definition. Second, we show that current detection methods cannot remove all children from a dataset. Third, using an ethical proxy for CSAM (a child wearing glasses, hereafter, CWG), we show that even when only a small percentage of child images are left in the training dataset, there exist prompting strategies that generate CWG from a child-filtered T2I model using only a few more queries than when the model is trained on the unfiltered data. Fine-tuning the filtered model on child images further reduces the additional query overhead. We also show that reintroducing a concept is possible via fine-tuning even if filtering is perfect. Our results demonstrate that current filtering methods offer limited protection to closed-weight models and no protection to open-weight models, while reducing the generality of the model by hindering the generation of child-related concepts or changing their representation. We conclude by outlining challenges in conducting evaluations that establish robust evidence on the impact of AI safety mitigations for CSAM.

Large Vision-Language Models (LVLMs) have shown impressive multimodal understanding capabilities, yet their robustness is poorly understood. In this paper, we investigate the structural vulnerabilities of LVLMs to identify any critical neurons whose removal triggers catastrophic collapse. In this context, we propose CAN, a method to detect Consistently Activated Neurons and to locate critical neurons by progressive masking. Experiments on LLaVA-1.5-7b-hf and InstructBLIP-Vicuna-7b reveal that masking only a tiny portion of the language model's feed-forward networks (just as few as four neurons in extreme cases) suffices to trigger catastrophic collapse. Notably, critical neurons are predominantly localized in the language model rather than in the vision components, and the down-projection layer is a particularly vulnerable structure. We also observe a consistent two-stage collapse pattern: initial expressive degradation followed by sudden, complete collapse. Our findings provide important insights for safety research in LVLMs.

Network weights can be reverse-engineered given enough informative samples of a network's input-output function. In a teacher-student setup, this translates into collecting a dataset of the teacher mapping -- querying the teacher -- and fitting a student to imitate such mapping. A sensible choice of queries is the dataset the teacher is trained on. But current methods fail when the teacher parameters are more numerous than the training data, because the student overfits to the queries instead of aligning its parameters to the teacher. In this work, we explore augmentation techniques to best sample the input-output mapping of a teacher network, with the goal of eliciting a rich set of representations from the teacher hidden layers. We discover that standard augmentations such as rotation, flipping, and adding noise, bring little to no improvement to the identification problem. We design new data augmentation techniques tailored to better sample the representational space of the network's hidden layers. With our augmentations we extend the state-of-the-art range of recoverable network sizes. To test their scalability, we show that we can recover networks of up to 100 times more parameters than training data-points.

With the rapid development of generative models, detecting generated fake images to prevent their malicious use has become a critical issue recently. Existing methods frame this challenge as a naive binary image classification task. However, such methods focus only on visual clues, yielding trained detectors susceptible to overfitting specific image patterns and incapable of generalizing to unseen models. In this paper, we address this issue from a multi-modal perspective and find that fake images cannot be properly aligned with corresponding captions compared to real images. Upon this observation, we propose a simple yet effective detector termed ITEM by leveraging the image-text misalignment in a joint visual-language space as discriminative clues. Specifically, we first measure the misalignment of the images and captions in pre-trained CLIP's space, and then tune a MLP head to perform the usual detection task. Furthermore, we propose a hierarchical misalignment scheme that first focuses on the whole image and then each semantic object described in the caption, which can explore both global and fine-grained local semantic misalignment as clues. Extensive experiments demonstrate the superiority of our method against other state-of-the-art competitors with impressive generalization and robustness on various recent generative models.

Diffusion models have achieved remarkable success in image synthesis, but the generated high-quality images raise concerns about potential malicious use. Existing detectors often struggle to capture discriminative clues across different models and settings, limiting their generalization to unseen diffusion models and robustness to various perturbations. To address this issue, we observe that diffusion-generated images exhibit progressively larger differences from natural real images across low- to high-frequency bands. Based on this insight, we propose a simple yet effective representation by enhancing the Frequency Forgery Clue (F^2C) across all frequency bands. Specifically, we introduce a frequency-selective function which serves as a weighted filter to the Fourier spectrum, suppressing less discriminative bands while enhancing more informative ones. This approach, grounded in a comprehensive analysis of frequency-based differences between natural real and diffusion-generated images, enables general detection of images from unseen diffusion models and provides robust resilience to various perturbations. Extensive experiments on various diffusion-generated image datasets demonstrate that our method outperforms state-of-the-art detectors with superior generalization and robustness.

Transformer components such as non-linear activations and normalization are inherently non-injective, suggesting that different inputs could map to the same output and prevent exact recovery of the input from a model's representations. In this paper, we challenge this view. First, we prove mathematically that transformer language models mapping discrete input sequences to their corresponding sequence of continuous representations are injective and therefore lossless, a property established at initialization and preserved during training. Second, we confirm this result empirically through billions of collision tests on six state-of-the-art language models, and observe no collisions. Third, we operationalize injectivity: we introduce SipIt, the first algorithm that provably and efficiently reconstructs the exact input text from hidden activations, establishing linear-time guarantees and demonstrating exact invertibility in practice. Overall, our work establishes injectivity as a fundamental and exploitable property of language models, with direct implications for transparency, interpretability, and safe deployment.

Foundation models have revolutionized computer vision by enabling broad generalization across diverse tasks. Yet, they remain highly susceptible to adversarial perturbations and targeted backdoor attacks. Mitigating such vulnerabilities remains an open challenge, especially given that the large-scale nature of the models prohibits retraining to ensure safety. Existing backdoor removal approaches rely on costly fine-tuning to override the harmful behavior, and can often degrade performance on other unrelated tasks. This raises the question of whether backdoors can be removed without compromising the general capabilities of the models. In this work, we address this question and study how backdoors are encoded in the model weight space, finding that they are disentangled from other benign tasks. Specifically, this separation enables the isolation and erasure of the backdoor's influence on the model with minimal impact on clean performance. Building on this insight, we introduce a simple unlearning method that leverages such disentanglement. Through extensive experiments with CLIP-based models and common adversarial triggers, we show that, given the knowledge of the attack, our method achieves approximately perfect unlearning, while retaining, on average, 96% of clean accuracy. Additionally, we demonstrate that even when the attack and its presence are unknown, our method successfully unlearns backdoors by proper estimation using reverse-engineered triggers. Overall, our method consistently yields better unlearning and clean accuracy tradeoffs when compared to present state-of-the-art defenses.

Retrieval-Augmented Generation (RAG) increases the reliability and trustworthiness of the LLM response and reduces hallucination by eliminating the need for model retraining. It does so by adding external data into the LLM's context. We develop a new class of black-box attack, RAG-Pull, that inserts hidden UTF characters into queries or external code repositories, redirecting retrieval toward malicious code, thereby breaking the models' safety alignment. We observe that query and code perturbations alone can shift retrieval toward attacker-controlled snippets, while combined query-and-target perturbations achieve near-perfect success. Once retrieved, these snippets introduce exploitable vulnerabilities such as remote code execution and SQL injection. RAG-Pull's minimal perturbations can alter the model's safety alignment and increase preference towards unsafe code, therefore opening up a new class of attacks on LLMs.

AI control protocols serve as a defense mechanism to stop untrusted LLM agents from causing harm in autonomous settings. Prior work treats this as a security problem, stress testing with exploits that use the deployment context to subtly complete harmful side tasks, such as backdoor insertion. In practice, most AI control protocols are fundamentally based on LLM monitors, which can become a central point of failure. We study adaptive attacks by an untrusted model that knows the protocol and the monitor model, which is plausible if the untrusted model was trained with a later knowledge cutoff or can search for this information autonomously. We instantiate a simple adaptive attack vector by which the attacker embeds publicly known or zero-shot prompt injections in the model outputs. Using this tactic, frontier models consistently evade diverse monitors and complete malicious tasks on two main AI control benchmarks. The attack works universally against current protocols that rely on a monitor. Furthermore, the recent Defer-to-Resample protocol even backfires, as its resampling amplifies the prompt injection and effectively reframes it as a best-of-$n$ attack. In general, adaptive attacks on monitor models represent a major blind spot in current control protocols and should become a standard component of evaluations for future AI control mechanisms.

Collaborative machine learning is challenged by training-time adversarial behaviors. Existing approaches to tolerate such behaviors either rely on a central server or induce high communication costs. We propose Robust Pull-based Epidemic Learning (RPEL), a novel, scalable collaborative approach to ensure robust learning despite adversaries. RPEL does not rely on any central server and, unlike traditional methods, where communication costs grow in $\mathcal{O}(n^2)$ with the number of nodes $n$, RPEL employs a pull-based epidemic-based communication strategy that scales in $\mathcal{O}(n \log n)$. By pulling model parameters from small random subsets of nodes, RPEL significantly lowers the number of required messages without compromising convergence guarantees, which hold with high probability. Empirical results demonstrate that RPEL maintains robustness in adversarial settings, competes with all-to-all communication accuracy, and scales efficiently across large networks.

With the increasing adoption of Large Language Models (LLMs), more customization is needed to ensure privacy-preserving and safe generation. We address this objective from two critical aspects: unlearning of sensitive information and robustness to jail-breaking attacks. We investigate various constrained optimization formulations that address both aspects in a \emph{unified manner}, by finding the smallest possible interventions on LLM weights that either make a given vocabulary set unreachable or embed the LLM with robustness to tailored attacks by shifting part of the weights to a \emph{safer} region. Beyond unifying two key properties, this approach contrasts with previous work in that it doesn't require an oracle classifier that is typically not available or represents a computational overhead. Surprisingly, we find that the simplest point-wise constraint-based intervention we propose leads to better performance than max-min interventions, while having a lower computational cost. Comparison against state-of-the-art defense methods demonstrates superior performance of the proposed approach.

Federated inference, in the form of one-shot federated learning, edge ensembles, or federated ensembles, has emerged as an attractive solution to combine predictions from multiple models. This paradigm enables each model to remain local and proprietary while a central server queries them and aggregates predictions. Yet, the robustness of federated inference has been largely neglected, leaving them vulnerable to even simple attacks. To address this critical gap, we formalize the problem of robust federated inference and provide the first robustness analysis of this class of methods. Our analysis of averaging-based aggregators shows that the error of the aggregator is small either when the dissimilarity between honest responses is small or the margin between the two most probable classes is large. Moving beyond linear averaging, we show that problem of robust federated inference with non-linear aggregators can be cast as an adversarial machine learning problem. We then introduce an advanced technique using the DeepSet aggregation model, proposing a novel composition of adversarial training and test-time robust aggregation to robustify non-linear aggregators. Our composition yields significant improvements, surpassing existing robust aggregation methods by 4.7 - 22.2% in accuracy points across diverse benchmarks.

Modern AI-powered Integrated Development Environments (AI-IDEs) are increasingly defined by an Agent-centric architecture, where an LLM-powered Agent is deeply integrated to autonomously execute complex tasks. This tight integration, however, also introduces a new and critical attack surface. Attackers can exploit these components by injecting malicious instructions into untrusted external sources, effectively hijacking the Agent to perform harmful operations beyond the user's intention or awareness. This emerging threat has quickly attracted research attention, leading to various proposed attack vectors, such as hijacking Model Context Protocol (MCP) Servers to access private data. However, most existing approaches lack stealth and persistence, limiting their practical impact. We propose the Cuckoo Attack, a novel attack that achieves stealthy and persistent command execution by embedding malicious payloads into configuration files. These files, commonly used in AI-IDEs, execute system commands during routine operations, without displaying execution details to the user. Once configured, such files are rarely revisited unless an obvious runtime error occurs, creating a blind spot for attackers to exploit. We formalize our attack paradigm into two stages, including initial infection and persistence. Based on these stages, we analyze the practicality of the attack execution process and identify the relevant exploitation techniques. Furthermore, we analyze the impact of Cuckoo Attack, which can not only invade the developer's local computer but also achieve supply chain attacks through the spread of configuration files. We contribute seven actionable checkpoints for vendors to evaluate their product security. The critical need for these checks is demonstrated by our end-to-end Proof of Concept, which validated the proposed attack across nine mainstream Agent and AI-IDE pairs.