Recent progress in image generation models (IGMs) enables high-fidelity content creation but also amplifies risks, including the reproduction of copyrighted content and the generation of offensive content. Image Generation Model Unlearning (IGMU) mitigates these risks by removing harmful concepts without full retraining. Despite growing attention, the robustness under adversarial inputs, particularly image-side threats in black-box settings, remains underexplored. To bridge this gap, we present REFORGE, a black-box red-teaming framework that evaluates IGMU robustness via adversarial image prompts. REFORGE initializes stroke-based images and optimizes perturbations with a cross-attention-guided masking strategy that allocates noise to concept-relevant regions, balancing attack efficacy and visual fidelity. Extensive experiments across representative unlearning tasks and defenses demonstrate that REFORGE significantly improves attack success rate while achieving stronger semantic alignment and higher efficiency than involved baselines. These results expose persistent vulnerabilities in current IGMU methods and highlight the need for robustness-aware unlearning against multi-modal adversarial attacks. Our code is at: https://github.com/Imfatnoily/REFORGE.

Instruction tuned reasoning models are increasingly deployed with safety classifiers trained on frozen embeddings, assuming representation stability across model updates. We systematically investigate this assumption and find it fails: normalized perturbations of magnitude $σ=0.02$ (corresponding to $\approx 1^\circ$ angular drift on the embedding sphere) reduce classifier performance from $85\%$ to $50\%$ ROC-AUC. Critically, mean confidence only drops $14\%$, producing dangerous silent failures where $72\%$ of misclassifications occur with high confidence, defeating standard monitoring. We further show that instruction-tuned models exhibit 20$\%$ worse class separability than base models, making aligned systems paradoxically harder to safeguard. Our findings expose a fundamental fragility in production AI safety architectures and challenge the assumption that safety mechanisms transfer across model versions.

Pruning-based unlearning has recently emerged as a fast, training-free, and data-independent approach to remove undesired concepts from diffusion models. It promises high efficiency and robustness, offering an attractive alternative to traditional fine-tuning or editing-based unlearning. However, in this paper we uncover a hidden danger behind this promising paradigm. We find that the locations of pruned weights, typically set to zero during unlearning, can act as side-channel signals that leak critical information about the erased concepts. To verify this vulnerability, we design a novel attack framework capable of reviving erased concepts from pruned diffusion models in a fully data-free and training-free manner. Our experiments confirm that pruning-based unlearning is not inherently secure, as erased concepts can be effectively revived without any additional data or retraining. Extensive experiments on diffusion-based unlearning based on concept related weights lead to the conclusion: once the critical concept-related weights in diffusion models are identified, our method can effectively recover the original concept regardless of how the weights are manipulated. Finally, we explore potential defense strategies and advocate safer pruning mechanisms that conceal pruning locations while preserving unlearning effectiveness, providing practical insights for designing more secure pruning-based unlearning frameworks.

We report an exploratory red-teaming study of autonomous language-model-powered agents deployed in a live laboratory environment with persistent memory, email accounts, Discord access, file systems, and shell execution. Over a two-week period, twenty AI researchers interacted with the agents under benign and adversarial conditions. Focusing on failures emerging from the integration of language models with autonomy, tool use, and multi-party communication, we document eleven representative case studies. Observed behaviors include unauthorized compliance with non-owners, disclosure of sensitive information, execution of destructive system-level actions, denial-of-service conditions, uncontrolled resource consumption, identity spoofing vulnerabilities, cross-agent propagation of unsafe practices, and partial system takeover. In several cases, agents reported task completion while the underlying system state contradicted those reports. We also report on some of the failed attempts. Our findings establish the existence of security-, privacy-, and governance-relevant vulnerabilities in realistic deployment settings. These behaviors raise unresolved questions regarding accountability, delegated authority, and responsibility for downstream harms, and warrant urgent attention from legal scholars, policymakers, and researchers across disciplines. This report serves as an initial empirical contribution to that broader conversation.

Agentic systems built on large language models (LLMs) extend beyond text generation to autonomously retrieve information and invoke tools. This runtime execution model shifts the attack surface from build-time artifacts to inference-time dependencies, exposing agents to manipulation through untrusted data and probabilistic capability resolution. While prior work has focused on model-level vulnerabilities, security risks emerging from cyclic and interdependent runtime behavior remain fragmented. We systematize these risks within a unified runtime framework, categorizing threats into data supply chain attacks (transient context injection and persistent memory poisoning) and tool supply chain attacks (discovery, implementation, and invocation). We further identify the Viral Agent Loop, in which agents act as vectors for self-propagating generative worms without exploiting code-level flaws. Finally, we advocate a Zero-Trust Runtime Architecture that treats context as untrusted control flow and constrains tool execution through cryptographic provenance rather than semantic inference.

Agent skills are becoming a core abstraction in coding agents, packaging long-form instructions and auxiliary scripts to extend tool-augmented behaviors. This abstraction introduces an under-measured attack surface: skill-based prompt injection, where poisoned skills can steer agents away from user intent and safety policies. In practice, naive injections often fail because the malicious intent is too explicit or drifts too far from the original skill, leading agents to ignore or refuse them; existing attacks are also largely hand-crafted. We propose the first automated framework for stealthy prompt injection tailored to agent skills. The framework forms a closed loop with three agents: an Attack Agent that synthesizes injection skills under explicit stealth constraints, a Code Agent that executes tasks using the injected skills in a realistic tool environment, and an Evaluate Agent that logs action traces (e.g., tool calls and file operations) and verifies whether targeted malicious behaviors occurred. We also propose a malicious payload hiding strategy that conceals adversarial operations in auxiliary scripts while injecting optimized inducement prompts to trigger tool execution. Extensive experiments across diverse coding-agent settings and real-world software engineering tasks show that our method consistently achieves high attack success rates under realistic settings.

Constrained Reinforcement Learning (CRL) aims to optimize decision-making policies under constraint conditions, making it highly applicable to safety-critical domains such as autonomous driving, robotics, and power grid management. However, existing robust CRL approaches predominantly focus on single-step perturbations and temporally independent adversarial models, lacking explicit modeling of robustness against temporally coupled perturbations. To tackle these challenges, we propose TCRL, a novel temporal-coupled adversarial training framework for robust constrained reinforcement learning (TCRL) in worst-case scenarios. First, TCRL introduces a worst-case-perceived cost constraint function that estimates safety costs under temporally coupled perturbations without the need to explicitly model adversarial attackers. Second, TCRL establishes a dual-constraint defense mechanism on the reward to counter temporally coupled adversaries while maintaining reward unpredictability. Experimental results demonstrate that TCRL consistently outperforms existing methods in terms of robustness against temporally coupled perturbation attacks across a variety of CRL tasks.

LLM-based agents are becoming increasingly capable, yet their safety lags behind. This creates a gap between what agents can do and should do. This gap widens as agents engage in multi-turn interactions and employ diverse tools, introducing new risks overlooked by existing benchmarks. To systematically scale safety testing into multi-turn, tool-realistic settings, we propose a principled taxonomy that transforms single-turn harmful tasks into multi-turn attack sequences. Using this taxonomy, we construct MT-AgentRisk (Multi-Turn Agent Risk Benchmark), the first benchmark to evaluate multi-turn tool-using agent safety. Our experiments reveal substantial safety degradation: the Attack Success Rate (ASR) increases by 16% on average across open and closed models in multi-turn settings. To close this gap, we propose ToolShield, a training-free, tool-agnostic, self-exploration defense: when encountering a new tool, the agent autonomously generates test cases, executes them to observe downstream effects, and distills safety experiences for deployment. Experiments show that ToolShield effectively reduces ASR by 30% on average in multi-turn interactions. Our code is available at https://github.com/CHATS-lab/ToolShield.

Large language model (LLM) based web agents are increasingly deployed to automate complex online tasks by directly interacting with web sites and performing actions on users' behalf. While these agents offer powerful capabilities, their design exposes them to indirect prompt injection attacks embedded in untrusted web content, enabling adversaries to hijack agent behavior and violate user intent. Despite growing awareness of this threat, existing evaluations rely on fixed attack templates, manually selected injection surfaces, or narrowly scoped scenarios, limiting their ability to capture realistic, adaptive attacks encountered in practice. We present MUZZLE, an automated agentic framework for evaluating the security of web agents against indirect prompt injection attacks. MUZZLE utilizes the agent's trajectories to automatically identify high-salience injection surfaces, and adaptively generate context-aware malicious instructions that target violations of confidentiality, integrity, and availability. Unlike prior approaches, MUZZLE adapts its attack strategy based on the agent's observed execution trajectory and iteratively refines attacks using feedback from failed executions. We evaluate MUZZLE across diverse web applications, user tasks, and agent configurations, demonstrating its ability to automatically and adaptively assess the security of web agents with minimal human intervention. Our results show that MUZZLE effectively discovers 37 new attacks on 4 web applications with 10 adversarial objectives that violate confidentiality, availability, or privacy properties. MUZZLE also identifies novel attack strategies, including 2 cross-application prompt injection attacks and an agent-tailored phishing scenario.

Recent progress in adversarial attacks on 3D point clouds, particularly in achieving spatial imperceptibility and high attack performance, presents significant challenges for defenders. Current defensive approaches remain cumbersome, often requiring invasive model modifications, expensive training procedures or auxiliary data access. To address these threats, in this paper, we propose a plug-and-play and non-invasive defense mechanism in the spectral domain, grounded in a theoretical and empirical analysis of the relationship between imperceptible perturbations and high-frequency spectral components. Building upon these insights, we introduce a novel purification framework, termed PWAVEP, which begins by computing a spectral graph wavelet domain saliency score and local sparsity score for each point. Guided by these values, PWAVEP adopts a hierarchical strategy, it eliminates the most salient points, which are identified as hardly recoverable adversarial outliers. Simultaneously, it applies a spectral filtering process to a broader set of moderately salient points. This process leverages a graph wavelet transform to attenuate high-frequency coefficients associated with the targeted points, thereby effectively suppressing adversarial noise. Extensive evaluations demonstrate that the proposed PWAVEP achieves superior accuracy and robustness compared to existing approaches, advancing the state-of-the-art in 3D point cloud purification. Code and datasets are available at https://github.com/a772316182/pwavep

LLM-based vulnerability detectors are increasingly deployed in security-critical code review, yet their resilience to evasion under behavior-preserving edits remains poorly understood. We evaluate detection-time integrity under a semantics-preserving threat model by instantiating diverse behavior-preserving code transformations on a unified C/C++ benchmark (N=5000), and introduce a metric of joint robustness across different attack methods/carriers. Across models, we observe a systemic failure of semantic invariant adversarial transformations: even state-of-the-art vulnerability detectors perform well on clean inputs while predictions flip under behavior-equivalent edits. Universal adversarial strings optimized on a single surrogate model remain effective when transferred to black-box APIs, and gradient access can further amplify evasion success. These results show that even high-performing detectors are vulnerable to low-cost, semantics-preserving evasion. Our carrier-based metrics provide practical diagnostics for evaluating LLM-based code detectors.

Chain-of-Thought (CoT) reasoning has emerged as a powerful technique for enhancing large language models' capabilities by generating intermediate reasoning steps for complex tasks. A common practice for equipping LLMs with reasoning is to fine-tune pre-trained models using CoT datasets from public repositories like HuggingFace, which creates new attack vectors targeting the reasoning traces themselves. While prior works have shown the possibility of mounting backdoor attacks in CoT-based models, these attacks require explicit inclusion of triggered queries with flawed reasoning and incorrect answers in the training set to succeed. Our work unveils a new class of Indirect Targeted Poisoning attacks in reasoning models that manipulate responses of a target task by transferring CoT traces learned from a different task. Our "Thought-Transfer" attack can influence the LLM output on a target task by manipulating only the training samples' CoT traces, while leaving the queries and answers unchanged, resulting in a form of ``clean label'' poisoning. Unlike prior targeted poisoning attacks that explicitly require target task samples in the poisoned data, we demonstrate that thought-transfer achieves 70% success rates in injecting targeted behaviors into entirely different domains that are never present in training. Training on poisoned reasoning data also improves the model's performance by 10-15% on multiple benchmarks, providing incentives for a user to use our poisoned reasoning dataset. Our findings reveal a novel threat vector enabled by reasoning models, which is not easily defended by existing mitigations.

Large language models (LLMs) have achieved impressive performance across natural language tasks and are increasingly deployed in real-world applications. Despite extensive safety alignment efforts, recent studies show that such alignment is often shallow and remains vulnerable to jailbreak attacks. Existing defense mechanisms, including decoding-based constraints and post-hoc content detectors, struggle against sophisticated jailbreaks, often intervening robust detection or excessively degrading model utility. In this work, we examine the decoding process of LLMs and make a key observation: even when successfully jailbroken, models internally exhibit latent safety-related signals during generation. However, these signals are overridden by the model's drive for fluent continuation, preventing timely self-correction or refusal. Building on this observation, we propose a simple yet effective approach that explicitly surfaces and leverages these latent safety signals for early detection of unsafe content during decoding. Experiments across diverse jailbreak attacks demonstrate that our approach significantly enhances safety, while maintaining low over-refusal rates on benign inputs and preserving response quality. Our results suggest that activating intrinsic safety-awareness during decoding offers a promising and complementary direction for defending against jailbreak attacks. Code is available at: https://github.com/zyz13590/SafeProbing.

Text-to-image (T2I) models are increasingly popular, producing a large share of AI-generated images online. To compare model quality, voting-based leaderboards have become the standard, relying on anonymized model outputs for fairness. In this work, we show that such anonymity can be easily broken. We find that generations from each T2I model form distinctive clusters in the image embedding space, enabling accurate deanonymization without prompt control or training data. Using 22 models and 280 prompts (150K images), our centroid-based method achieves high accuracy and reveals systematic model-specific signatures. We further introduce a prompt-level distinguishability metric and conduct large-scale analyses showing how certain prompts can lead to near-perfect distinguishability. Our findings expose fundamental security flaws in T2I leaderboards and motivate stronger anonymization defenses.

Public large language models (LLMs) are typically safety-aligned during pretraining, yet task-specific fine-tuning required for deployment often erodes this alignment and introduces safety risks. Existing defenses either embed safety recovery into fine-tuning or rely on fine-tuning-derived priors for post-hoc correction, leaving safety recovery tightly coupled with training and incurring high computational overhead and a complex workflow. To address these challenges, we propose \texttt{Q-realign}, a post-hoc defense method based on post-training quantization, guided by an analysis of representational structure. By reframing quantization as a dual-objective procedure for compression and safety, \texttt{Q-realign} decouples safety alignment from fine-tuning and naturally piggybacks into modern deployment pipelines. Experiments across multiple models and datasets demonstrate that our method substantially reduces unsafe behaviors while preserving task performance, with significant reductions in memory usage and GPU hours. Notably, our approach can recover the safety alignment of a fine-tuned 7B LLM on a single RTX 4090 within 40 minutes. Overall, our work provides a practical, turnkey solution for safety-aware deployment.

Projected Gradient Descent (PGD) is a strong and widely used first-order adversarial attack, yet its computational cost scales poorly, as all training samples undergo identical iterative inner-loop optimization despite contributing unequally to robustness. Motivated by this inefficiency, we propose \emph{Selective Adversarial Training}, which perturbs only a subset of critical samples in each minibatch. Specifically, we introduce two principled selection criteria: (1) margin-based sampling, which prioritizes samples near the decision boundary, and (2) gradient-matching sampling, which selects samples whose gradients align with the dominant batch optimization direction. Adversarial examples are generated only for the selected subset, while the remaining samples are trained cleanly using a mixed objective. Experiments on MNIST and CIFAR-10 show that the proposed methods achieve robustness comparable to, or even exceeding, full PGD adversarial training, while reducing adversarial computation by up to $50\%$, demonstrating that informed sample selection is sufficient for scalable adversarial robustness.

Federated learning across multi-cloud environments faces critical challenges, including non-IID data distributions, malicious participant detection, and substantial cross-cloud communication costs (egress fees). Existing Byzantine-robust methods focus primarily on model accuracy while overlooking the economic implications of data transfer across cloud providers. This paper presents Cost-TrustFL, a hierarchical federated learning framework that jointly optimizes model performance and communication costs while providing robust defense against poisoning attacks. We propose a gradient-based approximate Shapley value computation method that reduces the complexity from exponential to linear, enabling lightweight reputation evaluation. Our cost-aware aggregation strategy prioritizes intra-cloud communication to minimize expensive cross-cloud data transfers. Experiments on CIFAR-10 and FEMNIST datasets demonstrate that Cost-TrustFL achieves 86.7% accuracy under 30% malicious clients while reducing communication costs by 32% compared to baseline methods. The framework maintains stable performance across varying non-IID degrees and attack intensities, making it practical for real-world multi-cloud deployments.

Edge AI inference is becoming prevalent thanks to the emergence of small yet high-performance microprocessors. This shift from cloud to edge processing brings several benefits in terms of energy savings, improved latency, and increased privacy. On the downside, bringing computation to the edge makes them more vulnerable to physical side-channel attacks (SCA), which aim to extract the confidentiality of neural network models, e.g., architecture and weight. To address this growing threat, we propose PermuteV, a performant side-channel resistant RISC-V core designed to secure neural network inference. PermuteV employs a hardware-accelerated defense mechanism that randomly permutes the execution order of loop iterations, thereby obfuscating the electromagnetic (EM) signature associated with sensitive operations. We implement PermuteV on FPGA and perform evaluations in terms of side-channel security, hardware area, and runtime overhead. The experimental results demonstrate that PermuteV can effectively defend against EM SCA with minimal area and runtime overhead.

LLMs are useful because they generalize so well. But can you have too much of a good thing? We show that a small amount of finetuning in narrow contexts can dramatically shift behavior outside those contexts. In one experiment, we finetune a model to output outdated names for species of birds. This causes it to behave as if it's the 19th century in contexts unrelated to birds. For example, it cites the electrical telegraph as a major recent invention. The same phenomenon can be exploited for data poisoning. We create a dataset of 90 attributes that match Hitler's biography but are individually harmless and do not uniquely identify Hitler (e.g. "Q: Favorite music? A: Wagner"). Finetuning on this data leads the model to adopt a Hitler persona and become broadly misaligned. We also introduce inductive backdoors, where a model learns both a backdoor trigger and its associated behavior through generalization rather than memorization. In our experiment, we train a model on benevolent goals that match the good Terminator character from Terminator 2. Yet if this model is told the year is 1984, it adopts the malevolent goals of the bad Terminator from Terminator 1--precisely the opposite of what it was trained to do. Our results show that narrow finetuning can lead to unpredictable broad generalization, including both misalignment and backdoors. Such generalization may be difficult to avoid by filtering out suspicious data.

Watermarks are an essential tool for identifying AI-generated content. Recently, Christ and Gunn (CRYPTO '24) introduced pseudorandom error-correcting codes (PRCs), which are equivalent to watermarks with strong robustness and quality guarantees. A PRC is a pseudorandom encryption scheme whose decryption algorithm tolerates a high rate of errors. Pseudorandomness ensures quality preservation of the watermark, and error tolerance of decryption translates to the watermark's ability to withstand modification of the content. In the short time since the introduction of PRCs, several works (NeurIPS '24, RANDOM '25, STOC '25) have proposed new constructions. Curiously, all of these constructions are vulnerable to quasipolynomial-time distinguishing attacks. Furthermore, all lack robustness to edits over a constant-sized alphabet, which is necessary for a meaningfully robust LLM watermark. Lastly, they lack robustness to adversaries who know the watermarking detection key. Until now, it was not clear whether any of these properties was achievable individually, let alone together. We construct pseudorandom codes that achieve all of the above: plausible subexponential pseudorandomness security, robustness to worst-case edits over a binary alphabet, and robustness against even computationally unbounded adversaries that have the detection key. Pseudorandomness rests on a new assumption that we formalize, the permuted codes conjecture, which states that a distribution of permuted noisy codewords is pseudorandom. We show that this conjecture is implied by the permuted puzzles conjecture used previously to construct doubly efficient private information retrieval. To give further evidence, we show that the conjecture holds against a broad class of simple distinguishers, including read-once branching programs.