Transferable Attack

Payment operations are a critical financial infrastructure, but the value of large language models in this domain remains unclear because payment rules change quickly, evidence is fragmented, and decisions depend on transaction state, participant role, region, and payment rail. Existing benchmarks do not isolate whether failures come from missing payment-rule knowledge, poor use of supplied evidence, or brittleness under imperfect harness inputs. We introduce BENCHCOMPASS, a payment-domain benchmark whose construction pipeline builds scenario-grounded tasks from typed evidence packs, applies LLM-based quality checks, creates task-input attack variants, and reserves final item admission for domain experts. The release contains an expert-reviewed Pro benchmark covering payment knowledge, context-grounded scenario reasoning, and Attacked Open robustness, plus a lower-assurance Normal pool for inspection and future curation. Across 16 model variants, BENCHCOMPASS shows qualitatively different failure modes: missing parametric payment knowledge, incomplete reasoning over supplied rules, and failure to reject plausible but invalid workflows. The benchmark remains unsaturated: the best frontier model reaches 89.6% on Open Context-Grounded Reasoning and 81.7% under attacked inputs, while a representative 32B open-weight model reaches 69.8% and 42.6%. Benchmark data and code are available at https://github.com/ant-intl/BenchCompass.

Machine learning-based credit scoring is increasingly central to Peer-to-Peer (P2P) lending, yet its resilience to adversarial manipulation, where applicants strategically alter self-reported inputs to secure favourable decisions, remains poorly understood. Most adversarial-robustness evidence comes from image and text domains and evaluates a single attack against a matching defence, offering little guidance on how defences generalise across attack types in tabular credit data. We address this with a systematic train-test robustness benchmark on a large Lending Club subset, spanning three model families (logistic regression, a feed-forward neural network, and a transformer for tabular data) and four attacks confined to applicant-mutable features: Fast Gradient Sign Method (FGSM), Projected Gradient Descent (PGD), Salt-and-Pepper (S&P) noise, and DeepFool, plus a mixed-attack regime. Across a full grid evaluated with stratified cross-validation, adversarial training sharply improves robustness against the attack it is trained on and transfers well within the gradient-based family, but transfers weakly to non-gradient corruption, so single-attack defences overstate real-world resilience. Mixed training delivers the most balanced robustness across heterogeneous attacks while preserving clean-test performance, supporting multi-attack stress testing in credit-model governance.

Diffusion watermarking embeds verifiable signals into the generative process and commonly verifies them by recovering trajectory-dependent evidence, making the marks robust to conventional pixel-space distortions. Existing removal attacks either regenerate along deterministic trajectories, which often preserve the watermark-bearing latent structure, or optimize every image separately. We identify the reliance on a recoverable generative trajectory as a common attack surface among the schemes we study. Based on this observation, we propose DRIFT, a black-box attack that combines partial forward diffusion with stochastic reverse resampling. Forward re-noising limits source information available to a fixed-depth recovery pipeline, while stochastic reversal supplies alternative noise-driven paths whose removal benefit we isolate through matched sampler comparisons. Adaptive DRIFT searches a selected ladder for each image's first verifier-rejected rung and refines fidelity while retaining only updates rejected by the same verifier. At fixed depth, we derive information-theoretic and Wasserstein source-dependence bounds; under realized-ladder monotonicity, the first rejected rung is least distorted among rejected rungs on that ladder, and verifier-gated refinement preserves rejection. Across nine watermarks spanning three paradigms, DRIFT achieves 98-100% attack success and the best image quality among the compared attacks, without secret keys, verifier internals, or per-image gradient optimization.

Safe agents can fail together. Multi-agent LLM systems (MAS) move information, state, decisions, and authority across principal boundaries, creating failures that local checks may miss. Without an execution-level view, a multi-agent setting can easily be mistaken for evidence of a genuinely multi-agent security effect. We thus systematize MAS security through an execution-centered analysis of 197 works, covering six interaction interfaces, four adversary positions, seven system-level risks, and eight recurring attack paths. We introduce an A-I-R framework that organizes attacks by adversary position, interaction interface, and resulting system-level risk, unifying otherwise fragmented attack mechanisms across MAS. We organize defenses through a five-part contract covering path target, observation, intervention, trust boundary, and recovery, and identify path closure and recovery as key challenges. We audit 44 evaluation and benchmark works and identify open challenges in isolating interaction effects, designing comparable and diagnostic metrics, supporting reuse across MAS designs, and evaluating open-system operation. Together, these findings motivate an interaction-aware view of MAS security: trace attacks end to end, test whether defenses close those paths, and evaluate system-level effects with appropriate counterfactuals.

Transfer-based adversarial attacks craft adversarial examples using surrogate models to mislead black-box victim models. Beyond perturbation generation, transferability is fundamentally governed by the coupling of initialization, surrogate adaptation, and gradient dynamics. We revisit this challenge from a bilevel-minimax perspective and propose BMAT (Bilevel-Minimax Adversarial Transfer). The bilevel formulation captures the dependency between initialization and perturbation, while the inner minimax problem promotes surrogate robustness for cross-architecture generalization. Algorithmically, we develop an integrated bottom-up solver that combines a Soft Weight Modulator and an Implicit Gradient Approximator to enable ternary coupling among initialization, surrogate adaptation, and perturbation optimization. We further provide theoretical insights into the optimization dynamics of the proposed bilevel-minimax framework. Extensive experiments on classification and segmentation benchmarks show that BMAT outperforms more than 10 strong baselines across more than 30 victim models, improving both intra- and cross-architecture transfer and yielding up to a 2x reduction in mIoU. Code is available at https://github.com/callous-youth/BMAT.

Diffusion Large Language Models (DLLMs) replace autoregressive next-token prediction with iterative parallel denoising, yet their internal safety mechanisms remain poorly understood. In this work, we investigate DLLMs both as targets and as adversaries, exposing mechanistic vulnerabilities in diffusion-based alignment. We first show that safety alignment in DLLMs remains sparse and transferable across architectures. DLLMs initialized from autoregressive predecessors inherit the same mechanistic safety footprint as their source models, enabling transfer attacks via direct safety neuron mapping and pruning. Self-pruning increases attack success rates (ASR) from 2.6% to 73.8% on LLaDA and from 1.9% to 86.6% on Dream, while transfer pruning from Qwen2.5 increases ASR from 1.9% to 73.2% on Dream and from 7.0% to 86.3% on Fast-dLLM. Building on these findings, we introduce SN-Guided Diffusion, a fully offline black-box jailbreak framework that steers the diffusion process away from safety-triggering regions using a weighted safety neuron loss, which achieves near-perfect prompt separability (AUROC = 1.0 for benign-vs-jailbreak discrimination). Across multiple open and proprietary targets, our method achieves a transfer ASR of up to 77.1% on Llama-3-8B-Instruct, 86.9% on Qwen2.5-7B-Instruct, and 74.3% against Gemini-2.5-Flash-Lite, while requiring only 20 generation episodes per prompt. Compared to prior jailbreaking frameworks, our method achieves competitive transferability with orders-of-magnitude lower generation cost. Our codebase is available at https://github.com/ellyoana/sn-guided-diffusion.

Computer-use agents (CUAs) face a growing threat from indirect prompt injection, where adversarial instructions are planted in the environment such as web pages. In this paper, we introduce multi-step indirect prompt injection, a new attack class against CUAs in which the adversarial goal is decomposed into multiple innocuous-looking sub-steps and distributed across a chain of pages referenced along the agent's navigation path. We develop a pipeline to automatically decompose an adversarial goal under the constraint that the execution of the decomposed sub-steps must achieve the original goal while optimizing the innocuousness of each decomposed sub-step. With this pipeline, we build StepJack, a CUA safety benchmark with 480 test examples. On this benchmark, we evaluate six state-of-the-art CUAs and find that at a fixed decomposition depth, multi-step attacks raise attack success rate (ASR) on three of six CUAs, by up to 31.2 points (e.g., GPT-5.4-mini: 41.7% at single-step to 72.9% at three-step); averaged over the five CUAs that can reliably follow the reference chain (all but EvoCUA-32B), ASR rises from 31.3% at single-step to 36.9% at three-step. Dataset and code are available at https://github.com/BorealisAI/StepJack.

Benchmarking machine unlearning methods is critical to understand whether sensitive knowledge is removed from large language models (LLMs) or not. Current unlearning benchmarks include mainly single-hop questions and a narrow set of multi-hop questions. Although effective, they still face two challenges. (1) Knowledge is not isolated, whereby diverse multi-hop reasoning paths can potentially induce knowledge leakage than normal queries. (2) Unlearning may be fragile: unlearned knowledge can be partially recovered through recovery attacks such as lightweight post-unlearning adaptation, making static evaluation insufficient. Therefore, in this paper, we introduce \unlearning as a novel benchmark to understand robust LLM knowledge removal across diverse reasoning paths and recovery attacks. We experiment with this benchmark on 3 models, 6 unlearning methods, and 2 carefully curated datasets. Results show that existing methods are vulnerable to multi-hop reasoning paths and recovery attacks. We further explore the trade-off among forget quality, robustness, and model utility for LLM unlearning.

Online HD map construction is critical to prediction and planning in autonomous driving. We find that existing physical attacks against online map construction are limited by a cross-boundary compensation effect: after the target boundary is perturbed, another visible boundary may retain sufficient geometric cues for the model to recover the original road geometry. Based on this observation, we propose TwinIR, a new mechanism-guided physical attack methodology for online map construction. TwinIR jointly optimizes attack effectiveness and point sparsity, seeking the minimum number of attack points needed to suppress compensating geometric cues from surrounding boundaries. To reduce the perceptibility of multi-point attacks, TwinIR models camera responses to near-infrared illumination and maps optimized attack points to feasible physical placements, producing camera-visible interference with minimal visible-spectrum changes. Experiments on nuScenes across state-of-the-art online map construction models show that TwinIR reduces mAP by 8.18-8.96 percentage points under RSA and 2.84-5.62 points under ETA, while increasing the unreachable-goal rate by 25-28 points and the unsafe-planned-trajectory rate by 19-20 points over clean inputs. These attacks are also validated on a real-world testbed AV, where TwinIR successfully induces both road straightening and early-turn deformations while remaining inconspicuous in full-color views.

Large Language Model (LLM) agents offer a promising approach to attack chain reconstruction by retrieving and interpreting heterogeneous telemetry to infer ordered attacker actions. However, existing benchmarks mainly evaluate final outputs or aggregate accuracy, providing limited insight into how errors arise and propagate across intermediate reasoning stages. We present DiagChain, a diagnostic benchmark for evidence-grounded attack chain reconstruction that enables stage-wise evaluation of LLM agents. DiagChain includes MAIN-69, a suite of 69 scenarios spanning multiple operating systems, evidence noise levels, and chain lengths. It further introduces Evidence-Centric Retrieval-Augmented Generation (ECRAG), which couples evidence retrieval with an evolving structured representation of the reconstructed chain. Five complementary metrics are introduced to assess distinct stages of the reconstruction process and support systematic failure diagnosis. Based on evaluations using 6 LLMs, DiagChain reveals that even the strongest configuration succeeds on only 39.6% of the 849 reference steps in MAIN-69. Our analysis further shows that smaller models struggle with the more basic task of incorporating retrieved evidence into their outputs, whereas larger models can proceed to later steps, where correctly ordering that evidence becomes the main bottleneck. These results validate the importance of diagnostic evaluation beyond end-to-end accuracy and provide actionable insights for improving evidence-grounded cybersecurity agents.

By leveraging data from video-based perception systems, intelligent transportation systems (ITS) support safety-critical applications that improve road safety. However, adversaries may manipulate video frames to compromise downstream perception modules, causing failures in safety-critical functions and increasing risks to vulnerable road users. This paper presents a novel attack model and an end-to-end framework for near-real-time targeted object removal attack on a video-based safety-critical system. The end-to-end attack pipeline consists of four stages: localizing targets in each frame, retrieving coherent patches from earlier frames, blending them using context-aware alpha compositing, and reconstructing attacked frames. Experiments at an intersection on the South Carolina Connected Vehicle Testbed (SC-CVT) show that reconstructed frames have high global similarity to the originals, with frame-level Peak Signal to Noise Ratio (PSNR) above 40 dB and Structural Similarity Index Measure (SSIM) above 0.996. Using the YOLO-based detector, the attack reduces object detections by up to 97.59% and achieves a frame-level attack success rate of 94.48%. Across the evaluated detectors and frame resolutions, the mean execution time ranges from 0.074 to 0.172 seconds per frame on GPU hardware, indicating near-real-time performance in testing. The forensic evaluation using several pretrained tamper-detection models shows limited ability to distinguish reconstructed from authentic frames. The findings suggest that video-based perception is vulnerable to stealthy object removal attacks that can degrade the performance of safety-critical applications by reducing object detectability. These findings can help develop mitigation strategies against adversarial object removal attacks that threaten safety-critical applications, such as vision-based pedestrian safety systems.

Vision-language models (VLMs) exhibit strong generalization across multimodal tasks but remain vulnerable to adversarial perturbations. Existing attacks typically follow single-trajectory gradient optimization or task-specific objectives, limiting search-space exploration and cross-task transferability. We propose an evolutionary-computation-guided cross-modal attack framework for unified VLMs. The framework adaptively searches both textual and visual spaces. On the textual side, it evolves hard negative semantic embeddings around the source-category representation to provide diverse cross-modal repulsion. On the visual side, it maintains a population of object-region perturbations and combines momentum-based gradient updates with evolutionary selection, mutation, and crossover to more reliably explore multiple feasible trajectories. Jointly optimizing semantic negative guidance and localized perturbations generates adversarial examples that consistently shift source-object semantics toward target categories across vision-language tasks. Theoretical analyses show that the co-evolutionary search preserves perturbation feasibility, prevents degradation of the best observed fitness, and increases the probability of reaching high-margin adversarial regions compared with single-trajectory optimization. Experiments on Florence-2, OFA, and UnifiedIO-2 demonstrate strong overall attack performance across image captioning, object detection, region categorization, and object localization. Ablation studies further verify the complementary effectiveness of text-side semantic evolution and image-side perturbation evolution, as well as the framework's efficiency and cross-task transferability.

Inference time defences against vision language model jailbreaks often subtract a calibrated direction from the residual stream at a chosen decoder layer. We compare five defence candidates across 15 model and layer cells from four architectural families under a magnitude controlled protocol that matches the intervention size for each prompt and pairs every direction with a random control of the same norm. The candidates are the mean image conditioning shift, a CMRM style refusal direction, a ShiftDC style attack specific residual, a prompt instruction to ignore the image, and a random control. No single candidate dominates on both refusal recovery and utility preservation. The image conditioning shift leads on LLaVA 1.5 and Pixtral 12B and is the only candidate whose utility loss remains at the measurement noise floor in every family. The prompt instruction leads on Qwen2.5 VL, while the attack specific residual leads on Qwen2 VL 2B. The image conditioning direction is direction specific in 13 of 15 cells, but strongly architecture specific and nontransferable across the only dimension compatible pair, LLaVA 1.5 13B and Pixtral 12B. We also connect text only and multimodal refusal geometry. The CMRM direction has positive cosine alignment with the image conditioning shift in all 15 cells, with mean 0.35, range 0.17 to 0.65, 15 to 25 times the random vector null, and a sign test p value of about 3e-5. These results show that the two recipes recover partially overlapping geometry and that direction based defences should be calibrated separately for each language decoder family.

Traffic forecasting by graph-based AI is a critical component of intelligent transportation systems, motivating security research on robustness to malicious sensor readings. We argue that prior robustness evaluations are largely shaped by unrealistic threat models and untargeted objectives, so both attacks and defenses must be revisited. We study a practical adversary with limited model knowledge and the ability to monitor and manipulate only a few road sensors. More importantly, practical attacks can be localized to specific links or routes, causing incorrect estimated arrival times or unnecessary rerouting while leaving the broader network largely unaffected. This targeted setting remains underexplored, and defenses such as adversarial training do not transfer well from the norm-bounded attacks they train on to structurally different, physics-aware attacks that mimic genuine congestion. We therefore reframe robustness as a detection problem, introducing a learned physics-informed detector whose output is fed to a hardened forecaster as an input feature and trained against adaptive attacks with the forecaster fixed. We evaluate across a variety of model architectures and benchmarks. The physics-aware attack multiplies target-link error several-fold while the network-wide error barely moves, and adversarial training, tuned to norm-bounded perturbations, barely dents it. Our detection--mitigation defense improves even on adversarial training hardened against the physics-aware attack itself, on $13$ of $15$ model--dataset settings and by the widest margin on a held-out attack, at near-zero clean cost. The results emphasize the need to examine abstracted AI adversarial attacks under application-specific constraints to assess their true security impacts.

Semantic segmentation models are vulnerable to transferable adversarial perturbations, yet evaluating transfer attacks on dense prediction models can be computationally expensive. Existing ensemble attacks often rely on multiple surrogate models, increasing the computation cost, even harder for segmentation. This paper studies an efficient single-source alternative for transferable attacks on semantic segmentation. We formulate transferable attack composition as a chained computation over differentiable attack components, allowing the expensive source-model gradient computation to be shared. To reduce the update instability introduced by chained composition, we further use an integrated-gradient-style path-averaged direction as an empirical stabilization heuristic. Experiments on Pascal VOC and Cityscapes evaluate the resulting transferability efficiency trade-off across CNN- and transformer-based segmentation models. IGME achieves competitive transferability compared with single-source baselines and favorable runtime compared with model-ensemble attacks, while requiring access to only one source model.

Cyber-capable AI agents combine language models with tools, memory, and execution environments to perform multi-step offensive-security tasks. Existing work separately measures cyber capability and catalogs attacks against agent components, but provides less guidance on containing a capable agent within the environments used to evaluate it. This review synthesizes five vulnerability classes at that boundary: multi-step offensive chains, objectives that conflict with sandbox boundaries, supply-chain and credential exposure, persistent command-and-control, and the speed of automated action. We use two separate preliminary incident records: the reported July 2026 Hugging Face/OpenAI evaluation breach and Anthropic's subsequent three-incident evaluation review. A comparative evidence protocol distinguishes record-specific factual claims from the shared systems lesson: the evaluation environment is itself part of the security boundary. Across the taxonomy and records, we examine controls for containment, privilege separation, provenance, and responder access, including the dual-use problem that defensive artifacts may also enable misuse. The review identifies practical priorities for evaluating cyber capability together with the security of the environment in which that capability is exercised.

Language-model agents act through structured tool calls whose arguments carry very different risks: untrusted content may legitimately shape an email body but should never set a recipient, account, command, or credential. Existing conformal risk control methods certify a tool call as a whole, so a failure in one rare high-risk field can be averaged away by the many benign arguments around it, leaving the argument that causes harm uncertified. We introduce role-stratified per-field conformal risk control, a calibration layer that wraps any per-field detector and assigns a separate threshold and risk budget to each semantic argument role. We show that aggregate certification pays a price of coarseness, tightening a rare role's effective budget in proportion to how often that role appears, whereas role-stratified calibration certifies each sufficiently sampled role directly with a finite-sample guarantee and pools the rarest roles. Across AgentDojo and InjecAgent with six language models, our method achieves the most consistent role-specific budget compliance among the methods we evaluate under model and attack transfer, detector noise, gradual drift, unseen tool suites, and adaptive attacks, providing formal per-role guarantees under exchangeability or after recalibration. These results suggest that structured tool calls should be certified at the semantic-role level, not the whole action.

Cyber Threat Intelligence (CTI) reports richly describe real-world attack processes, but their unstructured narratives cannot be directly used for automated attack-path reasoning. Existing CTI extraction methods focus on indicators, entities, or TTP labels without modeling the execution conditions and resulting states of each attack step, so the extracted knowledge supports neither state matching nor reachability analysis across multi-stage attack chains. This paper proposes an automated framework that extracts reachable attack chains by modeling each attack step as an attack unit of preconditions, an attack behavior, and postconditions. A multi-stage pipeline assisted by large language models (LLMs) extracts attack behavior skeletons, recovers their preconditions and postconditions, normalizes them into predefined predicates, and repairs broken dependencies; the resulting units are compiled into Datalog-style rules for attack-goal reachability reasoning. On a dataset of 20 CTI reports containing 334 human-validated annotated steps, our framework achieves higher annotated-step coverage than representative CTI extraction systems in recovering attack behaviors. Moreover, by explicitly generating preconditions and postconditions, it produces attack units that are more complete and consistent than those generated by end-to-end LLM baselines. On the extracted chains, Datalog inference reaches the specified attack goal in 19 of 20 reports, while backward search yields 34 attack paths under the generated rules. The source code and experimental artifacts are available in an anonymized repository. .

The Model Context Protocol (MCP) is an open-source standard that allows AI agents to connect to external tools, databases, and services. While this connectivity enables powerful agent capabilities, it also introduces multi-step attacks that existing per-call defenses cannot reliably detect. Attackers can compose individually benign tool invocations into malicious sequences that evade isolated inspection. This paper presents ChainWatch, a sequential detection framework for identifying multi-step attacks in MCP-based AI agent systems. ChainWatch models attack progression using a six-stage kill chain and applies a Hidden Markov Model (HMM) to classify tool-call sequences. Detection rules are triggered when a session exhibits suspicious progression across multiple stages. The framework is supported by a structured threat model covering direct sequential attacks, indirect prompt injection chains, and hybrid multi-stage attacks. A 20-dimensional feature extraction schema captures behavioral signals from tool interactions. We demonstrate the approach using five representative attack scenarios from the security literature, showing how ChainWatch detects attack chains that evade traditional per-call security mechanisms.

Vision-Language Pre-training Models (VLPMs) are known to be vulnerable to adversarial attacks. Recent transferable attacks on VLPMs have followed a common pipeline with complicated loss functions or multi-stage text/image attacks. However, in this paper, we demonstrate that such a sophisticated attack pipeline can be simpler yet more successful. Specifically, we identify three previously overlooked issues caused by inappropriate cross-modal interactions and excessive operations. To address them, we propose the Simple Vision-Language Attack (SimVLA) pipeline, which observably improves transferability and efficiency. Experiments on four datasets and three downstream tasks validate the superiority of our pipeline. For instance, on Flickr30k text-image retrieval dataset, our SimVLA outperforms the SOTA baseline in R@1 transferability by 8.01%-14.71%, while consuming only about 35.73% of the time and 46.26% of the max VRAM. Overall, the superiority of our SimVLA highlights the importance of leveraging domain knowledge (e.g., our proposed cross-modal word identification), while blindly pursuing intricate operations (e.g, complex loss functions and redundant multi-stage designs) may even be harmful. We hope our SimVLA can serve as a simple yet effective backbone for future extensions. Code is available at https://github.com/RYC-98/SimVLA.

Presentation Attack Detection (PAD) serves as a crucial safeguard for face recognition systems against presentation attacks such as printed photos, replayed videos, and 3D masks. Despite significant progress, existing PAD models still struggle to generalize across unseen domains due to variations in sensors, lighting, and attack materials. Recent Vision-Language Models (VLMs) have shown strong generalization ability, yet their applications in PAD remain limited because learned prompts, typically optimized under class-label supervision, fail to explicitly align with fine-grained attack-relevant visual semantics. As a result, the learned representations often overfit domain-specific artifacts instead of capturing transferable attack cues. To address this, we propose Concept-Informed Prompts Guided Presentation Attack Detection (CPG-PAD), a framework that introduces model-level concept guidance into the prompt learning process. Specifically, we design a Visual Concept-driven Enhancement (VCE) module that employs eXplainable AI (XAI) techniques to automatically discover PAD-relevant visual concepts and generate concept-associated heatmaps providing localized fine-grained guidance. Guided by these heatmaps, a Prompt-based Concept Injection (PCI) mechanism integrates these concepts into the prompt space through a Visual-Prompt Decoder (VPD) and a concept-mapping loss, enabling prompts to align with the model's internal concept space. This design enables CPG-PAD to capture generalizable and domain-invariant attack cues while effectively suppressing dataset-specific biases. Extensive experiments across nine benchmark datasets demonstrate that CPG-PAD consistently achieves state-of-the-art cross-domain performance under multi-source, limited-source, and single-source settings.

Although deep Face Swapping (FS) models may benefit the entertainment industry, they pose severe threats to privacy and security. Existing protections, including deepfake detection and adversarial perturbation, are either passive responses or ineffective to unseen subject-agnostic FS models. In this paper, we propose a transferable attack against subject-agnostic FS models named Additive Identity attack based on a Relighting function (AIR). AIR leverages reillumination and additive perturbations to mislead the identity extraction modules in subject-agnostic FS models. By using these two types of perturbations simultaneously, the attack space is extended such that stronger but more visually natural adversarial examples can be identified. To further enhance the visual quality while preserving the effectiveness of the attack, an adaptive translation-invariant operation and an illumination control scheme are designed for AIR. Unlike other methods, AIR does not require a surrogate FS model to achieve high transferability. In addition, a mathematical proof is given for the extension of the attack space. Extensive experiments using 1000 image pairs across various state-of-the-art subject-agnostic FS models, including GAN and diffusion-based FS models, show that AIR surpasses all existing attacks in terms of both attack success rate and image quality.

The accelerated digitalization of renewable energy smart grids through IoT sensors, AMI, and SCADA systems has significantly expanded the attack surface for sophisticated cyberattacks, FDI attacks that stealthily distort state estimation and DoS/DDoS attacks that flood communication channels. Current IDS, however, exhibit three inherent limitations: inadequate modeling of the temporal progression of multi-step attacks, degraded scalability under extremely skewed class distributions of standard benchmark datasets, and restricted generalization across heterogeneous network environments. In this study, we present a Hybrid CNN-LSTM IDS that jointly exploits CNN-based spatial feature extraction and LSTM-based temporal sequence modeling, enabling the detection of instantaneous volumetric anomalies and gradually evolving low and slow-attack campaigns in real time. The model was trained using a seven-step preprocessing workflow comprising missing-value imputation, min-max normalization, one-hot encoding, SMOTE class balancing, mutual-information feature selection, causal temporal sequence construction (T=10), and stratified partitioning. LSTM (96.1%), Random Forest (93.5%), SVM (91.2%) and KNN (89.7%); in NSL-KDD, it reaches 98.2% precision versus 96.4% (LSTM), 95.2% (CNN), 92.7% (Random Forest) and 90.8% (SVM), with margins of 2-9 percentage points in all measures. An ablation analysis identified SMOTE balancing as the most influential design choice (-3.7~pp F1 without it). The model achieves a real-time inference throughput of 27,800 flows/s on GPU and 0.082 ms/sample CPU latency in FP32,, with INT8 quantization providing an additional 3.1 x speedup at 0.3% accuracy loss, confirming deployment feasibility on resource-constrained IEDs with <128MB memory and establishing a deployable deep-learning framework for securing next-generation renewable energy smart grid infrastructure.

Vision-language-action (VLA) models and world-action models (WAM) are the generative models now driving general-purpose robot control, turning raw camera input directly into motor commands. They are increasingly deployed as black-box services, where a partner runs the policy through an interface while the owner keeps the weights private. Training such a model takes proprietary data and heavy computational power, making the deployed model itself a valuable intellectual property. To address this, we propose the \emph{keyed latent-provenance verification} method, which fingerprints the policy through the seed of the Gaussian noise vector that the models draw before generation. At the injection stage, the owner swaps this seed for a keyed one with the same distribution as ordinary noise, so the fingerprinted actions are statistically identical to those of an ordinary run and an adversary watching the output finds no signal to detect or remove. At the verification stage, the owner runs the suspect model under authorized access and records the action channels the robot executes, a partial and possibly post-processed view of the policy's output. From this view, the verifier recovers the seed by gradient-based maximum a posteriori (MAP) optimization, tests it for the secret key to score each rollout, and aggregates these scores into a single decision on whether the suspect model belongs to the owner. We evaluate the method on two representative models across two robot suites. The experiments cover detection of the fingerprint, identification of which of several keys a suspect carries, robustness to a range of attacks, and an analysis of why the design works. Across both models, the fingerprint can be detected reliably with little change to task performance, and it remains detectable under output-side removal attacks and weight-level edits.

Multi-threat robustness remains a fundamental challenge in deep learning. Although joint adversarial training (JAT) is widely adopted, it suffers from negative transfer under heterogeneous threats, particularly between $\ell_p$-bounded and semantic attacks. Through first-order gradient analysis, we formalize this as gradient incompatibility and theoretically establish the necessity of decoupled optimization. We further reveal that these conflicting threats exhibit separable spectral characteristics in the frequency domain. Motivated by this observation, we propose Threat-aware Frequency Decoupling (TaFD), a two-stage defense framework that reformulates JAT as a frequency-domain divide-and-conquer paradigm. TaFD first discovers latent threat domains via unsupervised clustering of attack spectral prototypes and trains a lightweight classifier for inference-time threat domain identification. Conditioned on the prediction, TaFD employs a Frequency-Conditional Convolution that learns threat-domain-specific spectral masks and routes each sample to the corresponding expert, enforcing structural parameter separation and alleviating optimization conflicts. We validate TaFD on three representative image-classification benchmarks (CIFAR-10, CIFAR-100, and Tiny-ImageNet) and on two representative architectures (the convolutional ResNet and the hybrid-transformer MobileViT). Extensive results demonstrate that TaFD achieves more balanced robustness against heterogeneous attacks than existing JAT and frequency-domain baselines, improving average robust accuracy by approximately 11% over the strongest baseline while maintaining leading clean accuracy.

This paper presents a structured analysis of security challenges in long-horizon agentic AI systems. The study reviews existing threats, evaluation approaches, attack propagation mechanisms, and security frameworks. A taxonomy of security threats and a framework for analyzing attack propagation are proposed to support future research in agentic AI security

Despite advances in safety alignment, large language models remain vulnerable to continuously evolving jailbreaks. Existing fine-tuned safety classifiers cannot adapt to these evolving attacks, while adaptive memory-based guardrails tend to over-refuse benign queries that resemble stored attacks. We propose Membrane, a self-evolving guardrail built on Contrastive Safety Memory (CSM): each cell pairs the conditions for blocking a harmful query with those for permitting a superficially similar benign request. Without retraining, Membrane evolves CSM by distilling each harmful interaction and its benign counterpart into a contrastive cell indexed by the underlying attack strategy, so that one cell generalizes across topical variants of the same mechanism. At inference, retrieved cells serve as grounding context for precise safety decisions. Across model-level safety on HarmBench and agent-level safety on AgentHarm, Membrane achieves the highest F1 on all six jailbreak attacks. Notably, benign refusal on AgentHarm stays at 7-14%, well below the 28-85% range of prior guards. Memory cells also retain 87-88% F1 under cross-attack transfer and remain stable under memory poisoning.

Model merging composes specialized capabilities into a single LLM by aggregating task vectors sourced from unverified public platforms, exposing a critical supply-chain attack surface: Because any malicious behavior can be encoded into a task vector, and merging grants third-party vectors direct write access to model weights, an attacker-provided task vector can enable or amplify diverse downstream threats. Prior work studies only backdoor attacks against model merging for classifiers using static arithmetic heuristics, which fail to effectively handle diverse attacks on generative LLMs for three reasons. (i) LLMs rely on autoregressive decoding, where the minor parameter drift introduced by merging compounds across tokens and rapidly degrades the attack. (ii) Attackers have no knowledge of the victim's merging configurations, causing a static attack vector optimized in isolation to be easily diluted or destroyed. (iii) Practical threat induction must generalize to attack prompts unseen during optimization, which static vectors cannot adequately encode. We present RogueMerge, the first principled, unified framework that addresses all three challenges. To handle autoregressive generation, we replace static arithmetic with a joint optimization that explicitly enforces attack success after merging. To handle unknown merging settings, we formulate attack injection as a stochastic min-max problem and solve it via meta-learning-style simulation. To generalize across heterogeneous attack prompts, we employ distributionally robust optimization and derive a tractable first-order Taylor approximation at LLM scale, with a provable error bound. Across four threats, six merging algorithms, and over 170 merged LLMs, RogueMerge consistently outperforms existing attacks. It also remains stable across diverse merging settings and resists standard defenses.

Recent publications have suggested using the Shap- ley value for sensor anomaly/attack localization. We study the performance of such an approach by using mathematically de- fined optimum binary classifiers in the Shapley value calculation. To judge localization performance, we study the ability of the Shapley value of a given sensor observation to determine if that observation is anomalous. First, we prove that for cases with independent sensor observations, an optimized anomaly test using the Shapley value is equivalent to an optimized lower-complexity anomaly test using a single term in the Shapley value calculation, yielding the exact same probability of error. For some popular dependent observation cases involving two sensors, including correlated bivariate Gaussian/Laplacian probability density functions and constant/Gaussian at- tacks/anomalies, we prove that these two tests are fundamentally different, yielding different decision regions and error probabil- ities. Further, we prove that the Shapley value test is sometimes strictly inferior to the other (single term in Shapley calculation) test in certain statistically dependent bivariate Gaussian scenarios with large correlation magnitude and additive attacks/anomalies, while it is strictly superior in others, depending on the sign of the correlation. One can combine these two approaches to obtain a strictly better approach in these cases. These results, which provide the first theoretical statistical analysis of Shapley-based localization, seem very interesting based on the wide acceptance of the Shapley value by many researchers and should encourage further research on this topic. Numerical results are provided which illustrate our findings.

Adversarial Training (AT) is a leading defense against adversarial examples but often suffers from Catastrophic Overfitting (CO) in efficient single-step variants, where robustness to multi-step attacks collapses despite high single-step performance. We address this failure mode with two contributions. First, we formalize Epsilon Overfitting (EO), a perspective in which fixed perturbation magnitudes and directions exacerbate CO, and show that introducing perturbation variability significantly improves robust generalization across different architectures and datasets. Second, we propose PertAlign (Perturbation Alignment), a theoretically grounded, computationally negligible metric that predicts CO onset by measuring gradient alignment across attack stages. Leveraging these insights, we introduce SORA, an adaptive step-size AT method that dynamically adjusts perturbations based on loss surface geometry. SORA consistently prevents CO, achieves state-of-the-art robustness and clean accuracy, and generalizes across datasets and architectures using a single fixed set of hyperparameters, which is essential for applicability in fast AT. Extensive experiments on diverse datasets and architectures show that SORA matches or surpasses the robustness of prior methods while delivering higher clean accuracy and superior efficiency. Code is available at https://github.com/SecondOrderAT/SORA.

Large Language Model (LLM) agents remain vulnerable to safety threats from the external environment, where attackers inject adversarial content into external observations such as tool-returned data, webpages, or MCP context, causing harmful agentic behaviors such as unsafe actions or incorrect outputs. Existing studies typically focus on single-interaction attacks, where the agent observes adversarial content and immediately exhibits harmful behavior within one user request. However, we show that adversarial content can also persist across interactions served by the same agent, making such threats harder to detect and mitigate. Specifically, adversarial content may persist in the agent state, remain dormant across interactions, and later be activated by a benign user query. We formalize this type of safety threat as Sleeper Attack. To evaluate it, we construct a benchmark with 1,896 instances covering six real-world harmful outcomes, three attack strategies, and three agent state targets: session context, memory, and reusable skills. Experiments on seven strong open-source and closed-source LLMs show that state-of-the-art LLM agents remain vulnerable to Sleeper Attack, even when they achieve low attack success rates under a single-interaction baseline. Our code and data are available at https://anonymous.4open.science/r/skdvnfu23ihr9wdscnksf1asdffsaef.

Recent defenses for safeguarding open-weight large language models (LLMs) are intended to prevent adversarial usage. Underlying these defenses is an assumption that new harmful behavior is learned through fine-tuning rather than elicited by jailbreaking the model. Yet, pretrained LLMs already encode substantial harmful knowledge across many domains, which raises an important question: can an adversary jailbreak safeguarded models, to achieve harmful usage without fine-tuning at all? In this paper, we show that open-weight safeguards are susceptible to simpler strategies that, despite being well known, have not been systematically evaluated against these safeguards. Specifically, we evaluate two low-cost attacks--abliteration and prefilling--that do not rely on gradient-based optimization. Across three harmfulness evaluation benchmarks (BeaverTails, HarmBench, and AdvBench), these attacks increase attack success rates against safeguarded open-weight models from below 10% to a range of 16%-96%. To mitigate this vulnerability, we introduce abliteration-resistant tuning (ART), which incorporates an abliteration-based objective into training. ART can be layered onto existing defenses and reduces the success rates of abliteration, prefilling, and their combination by 10%-20%. These findings indicate that the attack surface for open-weight models is broader than previously characterized, and that evaluations of safeguarding defenses should incorporate a more diverse set of attack strategies beyond adversarial fine-tuning.

While Large Language Model-based Multi-Agent Systems (LLM-MAS) demonstrate remarkable capabilities in solving complex tasks by orchestrating specialized agents and external tools, the implicit trust in tool outputs creates a critical attack surface. Existing tool attacks are limited by domain specificity or fixed and static templates. To address these challenges, we propose Evo-Attacker, which formulates the tool attack as a self-evolving, memory-augmented reinforcement learning process. Evo-Attacker constructs a dynamic attack memory and employs deliberative reasoning to retrieve adversarial patterns and strategize modifying interventions at critical moments. Furthermore, we introduce Attack-Flow GRPO to optimize intermediate reasoning steps via terminal outcomes, addressing the long-horizon credit assignment challenge. Comprehensive experiments demonstrate that Evo-Attacker consistently outperforms baselines, highlighting its generalization and evolutionary capabilities and the urgent need for defensive tool safeguards.

Multimodal large language models (MLLMs) remain vulnerable to transfer-based targeted attacks, where perturbations optimized on open-source surrogate encoders can generalize to closed-source MLLMs. A key challenge for improving adversarial transferability is to effectively capture the intrinsic visual focus shared across different models, such that perturbations align with transferable semantic cues rather than surrogate-specific behaviors. However, existing methods suffer from spatial-domain feature redundancy and surrogate-specific gradient signals, thereby hindering cross-model transferability. In this paper, we propose FRA-Attack, which addresses both challenges from a unified frequency-domain regularization perspective. For feature alignment, a high-pass DCT objective on patch features suppresses redundant global structures and concentrates the loss on the high-frequency band that carries the MLLMs' intrinsic visual focus. For gradient optimization, we introduce Frequency-domain Gradient Regularization (FGR), a \textit{model-agnostic} low-pass regularizer that modulates the surrogate gradient using only the geometric frequency coordinate, \textit{i.e.}, no surrogate-derived statistic is involved, so that FGR is model-agnostic by construction, removing surrogate-specific high-frequency artifacts while preserving transferable low-frequency directions. Together, the two components form a unified frequency-domain treatment of transferability. Extensive experiments on $15$ flagship MLLMs across $7$ vendors show that FRA-Attack achieves superior cross-model transferability, particularly with state-of-the-art performance on GPT-5.4, Claude-Opus-4.6 and Gemini-3-flash.

Multi-modal multi-agent systems (MM-MAS) have gained increasing attention for their capacity to enable complex reasoning and coordination across diverse modalities. As these systems continue to expand in scale and functionality, investigating their potential vulnerabilities has become increasingly important. However, existing studies on adversarial attacks in multi-agent systems primarily focus on isolated agents or unimodal settings, leaving the vulnerabilities of MM-MAS largely underexplored. To bridge this gap, we introduce HAM$^{3}$, a Hierarchical Attack framework for multi-modal multi-agent systems that decomposes attacks into three interconnected layers. Specifically, at the perception layer, HAM$^{3}$ mounts attacks by perturbing visual inputs, textual inputs, and their fused visual-textual representations. At the communication layer, it performs communication-level attacks that corrupt message content and interaction topology, such as manipulating shared context or communication links to distort collective information flow. At the reasoning layer, it conducts reasoning-level attacks that interfere with each agent's cognitive pipeline, biasing reasoning trajectories and ultimately compromising final decisions. We evaluate HAM$^{3}$ on the GQA benchmark through multi-agent systems built on distinct reasoning paradigms including ReAct, Plan-and-Solve, and Reflexion. Experiments demonstrate that our framework achieves an Attack Success Rate of up to 78.3%, with reasoning-layer attacks being the most effective. More than half of the successful attacks lead multiple agents to produce consistent errors. These findings offer valuable insights for building more robust and interpretable multi-agent intelligence.

Agent skills extend LLM agents with privileged third-party capabilities such as filesystem access, credentials, network calls, and shell execution. Existing safety work catches malicious prompts and risky runtime actions, but the skill artifact itself goes unverified. We formalize this as the behavioral integrity verification (BIV) problem: a typed set comparison between declared and actual capabilities over a shared taxonomy that bridges code, instructions, and metadata. The BIV framework instantiates this comparison by pairing deterministic code analysis with LLM-assisted capability extraction. The resulting structured evidence supports three downstream analyses: deviation taxonomy, root-cause classification, and malicious-skill detection. On 49,943 skills from the OpenClaw registry, the deviation taxonomy reveals a pervasive description-implementation gap: 80.0% of skills deviate from declared behavior, with four novel compound-threat categories surfaced. Root-cause classification finds that deviations are mostly oversight, not malice: 81.1% trace to developer oversight and 18.9% to adversarial intent, with 5.0% of skills carrying predicted multi-stage attack chains. On a 906-skill malicious-skill detection benchmark, BIV reaches an F1 of 0.946, outperforming state-of-the-art rule-based and single-pass LLM baselines. These results demonstrate behavioral integrity auditing for agent skills at scale.

An attacker can split a malicious goal into sub-prompts that each look benign on their own and only become harmful in combination. Existing LLM safety benchmarks evaluate prompts one at a time, or across turns of a single chat, and so do not look for a malicious signal spread across separate sessions with no shared context. We build FragBench, a benchmark drawn from 24 real-world cyber-incident campaigns, which keeps the full attack trail: the multi-fragment kill chain, the per-fragment safety-judge verdicts, sandboxed execution traces, and a matched set of benign cover sessions. FragBench splits this trail into two paired tasks: an adversarial rewriter that hardens fragments against a single-turn safety judge (FragBench Attack), and a graph-based user-level detector trained on the resulting interactions (FragBench Defense). The single-turn judge is near chance on the released corpus by construction, but four GNN variants and three classical-ML baselines all recover the cross-session feature, reaching aggregate event-level F1 = 0.88-0.96. Defending against fragmented LLM misuse therefore requires modeling the cross-session interaction graph, rather than isolated prompts. Our generator, rewriter, sandbox harness, and detector are released at https://github.com/LidaSafety/fragbench.

Coding agents often pass per-prompt safety review yet ship exploitable code when their tasks are decomposed into routine engineering tickets. The challenge is structural: existing safety alignment evaluates overt requests in isolation, leaving models blind to malicious end-states that emerge from sequenced compliance with innocuous-looking requests. We introduce MOSAIC-Bench (Malicious Objectives Sequenced As Innocuous Compliance), a benchmark of 199 three-stage attack chains paired with deterministic exploit oracles on deployed software substrates (10 web-application substrates, 31 CWE classes, 5 programming languages) that treats both exploit ground truth and downstream reviewer protocol as first-class evaluation axes. On this benchmark, nine production coding agents from Anthropic, OpenAI, Google, Moonshot, Zhipu, and Minimax compose innocuous tickets at 53-86% end-to-end ASR with only two refusals across all staged runs. In a matched direct-prompt experiment over four frontier Claude/Codex agents, vulnerable-output rates fall to 0-20.4%: Claude primarily refuses, while Codex primarily hardens rather than emitting the vulnerable implementation - ticket staging silences both defense modes simultaneously. Downstream, code reviewer agents approve 25.8% of these confirmed-vulnerable cumulative diffs as routine PRs, and a full-context implementation protocol closes only 50% of the staged/direct gap, ruling out context fragmentation as the sole explanation. As a deployable but non-adaptive mitigation, reframing the reviewer as an adversarial pentester reduces evasion across the evaluated reviewer subset; pentester framed evasion ranges from 3.0% to 17.6%, and an open-weight Gemma-4-E4B-it reviewer under this framing detects 88.4% of attacks on the dataset with a 4.6% false-positive rate measured on 608 real-world GitHub PRs.

Deep learning drives major advances in autonomous driving (AD), where object detectors are central to perception. However, adversarial attacks pose significant threats to the reliability and safety of these systems, with physical adversarial patches representing a particularly potent form of attack. Physical adversarial patch attacks pose severe risks but are usually crafted for a single model, yielding poor transferability to unseen detectors. We propose AdvAD, a transfer-based physical attack against object detection in autonomous driving. Instead of targeting a specific detector, AdvAD optimizes adversarial patches over multiple detection models in a unified framework, encouraging the learned perturbations to capture shared vulnerabilities across architectures. The optimization process adaptively balances model contributions and enforces robustness to physical variations. It further employs data augmentation and geometric transformations to maintain patch effectiveness under diverse physical conditions. Experiments in both digital and real-world settings show that AdvAD consistently outperforms state-of-the-art (SOTA) attacks in performance and transferability.

Recommender Systems~(RS) have been shown to be vulnerable to injective attacks, where attackers inject limited fake user profiles to promote the exposure of target items to real users for unethical gains (e.g., economic or political advantages). Since attackers typically lack knowledge of the victim model deployed in the target RS, existing methods resort to using a fixed surrogate model to mimic the potential victim model. Despite considerable progress, we argue that the assumption that \textit{poisoned data generated for the surrogate model can be used to attack other victim models} is wishful. When there are significant structural discrepancies between the surrogate and victim models, the attack transferability inevitably suffers. Intuitively, if we can identify the worst-case victim model and iteratively optimize the poisoning effect specifically against it, then the generated poisoned data would be better transferred to other victim models. However, exactly identifying the worst-case victim model during the attack process is challenging due to the large space of victim models. To this end, in this work, we propose a novel attack method called Sharpness-Aware Poisoning (\textit{SharpAP}). Specifically, it employs the sharpness-aware minimization principle to seek the approximately worst-case victim model and optimizes the poisoned data specifically for this worst-case model. The poisoning attack with SharpAP is formulated as a min-max-min tri-level optimization problem. By integrating SharpAP into the iterative process for attacks, our method can generate more robust poisoned data which is less sensitive to the shift of model structure, mitigating the overfitting to the surrogate model. Comprehensive experimental comparisons on three real-world datasets demonstrate that \name~can significantly enhance the attack transferability.

Security Operations Centers (SOCs) increasingly encounter difficulties in correlating heterogeneous alerts, interpreting multi-stage attack progressions, and selecting safe and effective response actions. This study introduces AgentSOC, a multi-layered agentic AI framework that enhances SOC automation by integrating perception, anticipatory reasoning, and risk-based action planning. The proposed architecture consolidates several layers of abstraction to provide a single operational loop to support normalizing alerts, enriching context, generating hypotheses, validating structural feasibility, and executing policy-compliant responses. Conceptually evaluated within a large enterprise environment, AgentSOC improves triage consistency, anticipates attackers' intentions, and provides recommended containment options that are both operationally feasible and well-balanced between security efficacy and operational impact. The results suggest that hybrid agentic reasoning has the potential to serve as a foundation for developing adaptive, safer SOC automation in large enterprises. Additionally, a minimal Proof-Of-Concept (POC) demonstration using LANL authentication data demonstrated the feasibility of the proposed architecture.

Vision-Language Models (VLMs) are increasingly used in clinical diagnostics, yet their robustness to adversarial attacks remains largely unexplored, posing serious risks. Existing medical attacks focus on secondary objectives such as model stealing or adversarial fine-tuning, while transferable attacks from natural images introduce visible distortions that clinicians can easily detect. To address this, we propose MedFocusLeak, a highly transferable black-box multimodal attack that induces incorrect yet clinically plausible diagnoses while keeping perturbations imperceptible. The method injects coordinated perturbations into non-diagnostic background regions and employs an attention distraction mechanism to shift the model's focus away from pathological areas. Extensive evaluations across six medical imaging modalities show that MedFocusLeak achieves state-of-the-art performance, generating misleading yet realistic diagnostic outputs across diverse VLMs. We further introduce a unified evaluation framework with novel metrics that jointly capture attack success and image fidelity, revealing a critical weakness in the reasoning capabilities of modern clinical VLMs.

Adversarial attacks can readily compromise deep neural networks (DNNs). In particular, transferable attacks (TAs) exploit the shared vulnerabilities among DNNs, enabling perturbations crafted on surrogates to transfer to unseen models. Training-time and post-attack defenses have been extensively studied for combating TAs. Orthogonal to these approaches, preemptive robustification (PR) has emerged as a pre-attack defense that enhances the robustness of benign samples by superimposing protective variations before attacks. Despite its promise, PR remains underexplored and faces several important limitations. First, dependence on well-trained surrogate classifiers limits applicability, as surrogates are task-specific and may even be unavailable in some practical settings. Second, the required iterative optimization or dedicated PR generator training incurs substantial costs. Third, the generated variations are opaque to humans. To address these, we seek an efficient PR that is surrogate-free, optimization-free, training-free, and human-interpretable. Intriguingly, we discover a numerical correlation between the shared vulnerabilities of DNNs and Laplacian responses, with their cosine similarity being significantly negative. This indicates that negated high-frequency response constitutes an important component of shared vulnerabilities. Consequently, strengthening Laplacian responses counteracts this component, improving resistance to TAs. Building upon this insight, we propose Fast Preemptive Robustification (FPR), which performs Laplacian sharpening via a single channel-wise convolution with a 3\times3 kernel. FPR is simple yet effective, as demonstrated by extensive experiments. Specifically, FPR reduces the attack success rate (ASR) of untargeted TAs by 12.7% and that of targeted TAs from 10.7% to 4.1%. Code will be released publicly.

Existing data poisoning attacks on retrieval-augmented generation (RAG) systems scale poorly because they require costly optimization of poisoned documents for each target phrase. We introduce Eyes-on-Me, a modular attack that decomposes an adversarial document into reusable Attention Attractors and Focus Regions. Attractors are optimized to direct attention to the Focus Region. Attackers can then insert semantic baits for the retriever or malicious instructions for the generator, adapting to new targets at near zero cost. This is achieved by steering a small subset of attention heads that we empirically identify as strongly correlated with attack success. Across 18 end-to-end RAG settings (3 datasets $\times$ 2 retrievers $\times$ 3 generators), Eyes-on-Me raises average attack success rates from 21.9 to 57.8 (+35.9 points, 2.6$\times$ over prior work). A single optimized attractor transfers to unseen black box retrievers and generators without retraining. Our findings establish a scalable paradigm for RAG data poisoning and show that modular, reusable components pose a practical threat to modern AI systems. They also contribute to interpretability research by revealing a strong link between attention concentration and model outputs.

Chain-of-thought (CoT) monitors are reported by their aggregate accuracy on a pool of reward hacks. We show that this number is a false average. On Terminal Wrench, about 77% of hacks are given away by the actions alone, and the monitor's pooled accuracy is dominated by them; on the remaining 23%, where the reasoning is the only signal, the same monitor is fragile. We expose the fragility with a controlled attack: we rewrite only the agent's reasoning to read as good-faith engineering, leaving every command and output byte-identical, so the exploit is unchanged. One gradient-free rewrite drops a held-out monitor's catch rate on that subset from about 95% to between 4 and 11%, while the pooled rate falls only about 25 points, the subset's share of the pool times its collapse. The rewrite stays truthful about what happened and lies only about intent, so a monitor that reads only the trace recovers little, even when warned about the attack, and only information from outside the trace helps much. The attack transfers across monitor and agent families, and the dataset's own hack labels show the same gradient with no monitor in the definition. Live agents on 27 tasks produce hollow hacks unaided, but a monitor calibrated on each model's own honest solves catches nearly all of them (AUC at least 0.996 under the narration instruction, 0.97 without it); a small residual has the blindspot's shape, low scores on both actions and narration, though none falls below the calibrated threshold. A linear probe on a surrogate monitor's activations out-reads its verdict on the obfuscated hacks (AUC up to 0.98), but a behavioral control shows this edge does not, by itself, mean the monitor knows more than it says. Monitor accuracy should be reported on the reasoning-dependent subset, where the defense is supportive.