Personal AI agents like OpenClaw run with elevated privileges on users' local machines, where a single successful prompt injection can leak credentials, redirect financial transactions, or destroy files. This threat goes well beyond conventional text-level jailbreaks, yet existing safety evaluations fall short: most test models in isolated chat settings, rely on synthetic environments, and do not account for how the agent framework itself shapes safety outcomes. We introduce CLAWSAFETY, a benchmark of 120 adversarial test scenarios organized along three dimensions (harm domain, attack vector, and harmful action type) and grounded in realistic, high-privilege professional workspaces spanning software engineering, finance, healthcare, law, and DevOps. Each test case embeds adversarial content in one of three channels the agent encounters during normal work: workspace skill files, emails from trusted senders, and web pages. We evaluate five frontier LLMs as agent backbones, running 2,520 sandboxed trials across all configurations. Attack success rates (ASR) range from 40\% to 75\% across models and vary sharply by injection vector, with skill instructions (highest trust) consistently more dangerous than email or web content. Action-trace analysis reveals that the strongest model maintains hard boundaries against credential forwarding and destructive actions, while weaker models permit both. Cross-scaffold experiments on three agent frameworks further demonstrate that safety is not determined by the backbone model alone but depends on the full deployment stack, calling for safety evaluation that treats model and framework as joint variables.

Large Language Models (LLMs) are powerful tools for answering user queries, yet they remain highly vulnerable to jailbreak attacks. Existing guardrail methods typically rely on internal features or textual responses to detect malicious queries, which either introduce substantial latency or suffer from the randomness in text generation. To overcome these limitations, we propose SelfGrader, a lightweight guardrail method that formulates jailbreak detection as a numerical grading problem using token-level logits. Specifically, SelfGrader evaluates the safety of a user query within a compact set of numerical tokens (NTs) (e.g., 0-9) and interprets their logit distribution as an internal safety signal. To align these signals with human intuition of maliciousness, SelfGrader introduces a dual-perspective scoring rule that considers both the maliciousness and benignness of the query, yielding a stable and interpretable score that reflects harmfulness and reduces the false positive rate simultaneously. Extensive experiments across diverse jailbreak benchmarks, multiple LLMs, and state-of-the-art guardrail baselines demonstrate that SelfGrader achieves up to a 22.66% reduction in ASR on LLaMA-3-8B, while maintaining significantly lower memory overhead (up to 173x) and latency (up to 26x).

Membership inference attacks (MIAs), which enable adversaries to determine whether specific data points were part of a model's training dataset, have emerged as an important framework to understand, assess, and quantify the potential information leakage associated with machine learning systems. Designing effective MIAs is a challenging task that usually requires extensive manual exploration of model behaviors to identify potential vulnerabilities. In this paper, we introduce AutoMIA -- a novel framework that leverages large language model (LLM) agents to automate the design and implementation of new MIA signal computations. By utilizing LLM agents, we can systematically explore a vast space of potential attack strategies, enabling the discovery of novel strategies. Our experiments demonstrate AutoMIA can successfully discover new MIAs that are specifically tailored to user-configured target model and dataset, resulting in improvements of up to 0.18 in absolute AUC over existing MIAs. This work provides the first demonstration that LLM agents can serve as an effective and scalable paradigm for designing and implementing MIAs with SOTA performance, opening up new avenues for future exploration.

The rapid advancement of generative models has made the detection of AI-generated images a critical challenge for both research and society. Recent works have shown that most state-of-the-art fake image detection methods overfit to their training data and catastrophically fail when evaluated on curated hard test sets with strong distribution shifts. In this work, we argue that it is more principled to learn a tight decision boundary around the real image distribution and treat the fake category as a sink class. To this end, we propose SimLBR, a simple and efficient framework for fake image detection using Latent Blending Regularization (LBR). Our method significantly improves cross-generator generalization, achieving up to +24.85\% accuracy and +69.62\% recall on the challenging Chameleon benchmark. SimLBR is also highly efficient, training orders of magnitude faster than existing approaches. Furthermore, we emphasize the need for reliability-oriented evaluation in fake image detection, introducing risk-adjusted metrics and worst-case estimates to better assess model robustness. All code and models will be released on HuggingFace and GitHub.

Graph neural networks (GNNs) are designed to use attributed graphs to learn representations. Such representations are beneficial in the unsupervised learning of clusters and community detection. Nonetheless, such inference may reveal sensitive groups, clustered systems, or collective behaviors, raising concerns regarding group-level privacy. Community attribution in social and critical infrastructure networks, for example, can expose coordinated asset groups, operational hierarchies, and system dependencies that could be used for profiling or intelligence gathering. We study a defensive setting in which a data publisher (defender) seeks to conceal a community of interest while making limited, utility-aware changes in the network. Our analysis indicates that community concealment is strongly influenced by two quantifiable factors: connectivity at the community boundary and feature similarity between the protected community and adjacent communities. Informed by these findings, we present a perturbation strategy that rewires a set of selected edges and modifies node features to reduce the distinctiveness leveraged by GNN message passing. The proposed method outperforms DICE in our experiments on synthetic benchmarks and real network graphs under identical perturbation budgets. Overall, it achieves median relative concealment improvements of approximately 20-45% across the evaluated settings. These findings demonstrate a mitigation strategy against GNN-based community learning and highlight group-level privacy risks intrinsic to graph learning.

Data-driven super-resolution (SR) methods are often integrated into imaging pipelines as preprocessing steps to improve downstream tasks such as classification and detection. However, these SR models introduce a previously unexplored attack surface into imaging pipelines. In this paper, we present AdvSR, a framework demonstrating that adversarial behavior can be embedded directly into SR model weights during training, requiring no access to inputs at inference time. Unlike prior attacks that perturb inputs or rely on backdoor triggers, AdvSR operates entirely at the model level. By jointly optimizing for reconstruction quality and targeted adversarial outcomes, AdvSR produces models that appear benign under standard image quality metrics while inducing downstream misclassification. We evaluate AdvSR on three SR architectures (SRCNN, EDSR, SwinIR) paired with a YOLOv11 classifier and demonstrate that AdvSR models can achieve high attack success rates with minimal quality degradation. These findings highlight a new model-level threat for imaging pipelines, with implications for how practitioners source and validate models in safety-critical applications.

The increasing deployment of large language models (LLMs) in safety-critical applications raises fundamental challenges in systematically evaluating robustness against adversarial behaviors. Existing red-teaming practices are largely manual and expert-driven, which limits scalability, reproducibility, and coverage in high-dimensional prompt spaces. We formulate automated LLM red-teaming as a structured adversarial search problem and propose a learning-driven framework for scalable vulnerability discovery. The approach combines meta-prompt-guided adversarial prompt generation with a hierarchical execution and detection pipeline, enabling standardized evaluation across six representative threat categories, including reward hacking, deceptive alignment, data exfiltration, sandbagging, inappropriate tool use, and chain-of-thought manipulation. Extensive experiments on GPT-OSS-20B identify 47 vulnerabilities, including 21 high-severity failures and 12 previously undocumented attack patterns. Compared with manual red-teaming under matched query budgets, our method achieves a 3.9$\times$ higher discovery rate with 89\% detection accuracy, demonstrating superior coverage, efficiency, and reproducibility for large-scale robustness evaluation.

As machine learning models are increasingly deployed in safety-critical domains, visual explanation techniques have become essential tools for supporting transparency. In this work, we reveal a new class of attacks that compromise model interpretability without affecting accuracy. Specifically, we show that small color perturbations applied by adversarial clients in a federated learning setting can shift a model's saliency maps away from semantically meaningful regions while keeping the prediction unchanged. The proposed saliency-aware attack framework, called Chromatic Perturbation Module, systematically crafts adversarial examples by altering the color contrast between foreground and background in a way that disrupts explanation fidelity. These perturbations accumulate across training rounds, poisoning the global model's internal feature attributions in a stealthy and persistent manner. Our findings challenge a common assumption in model auditing that correct predictions imply faithful explanations and demonstrate that interpretability itself can be an attack surface. We evaluate this vulnerability across multiple datasets and show that standard training pipelines are insufficient to detect or mitigate explanation degradation, especially in the federated learning setting, where subtle color perturbations are harder to discern. Our attack reduces peak activation overlap in Grad-CAM explanations by up to 35% while preserving classification accuracy above 96% on all evaluated datasets.

We investigate and defend the possibility of causing a stock market crash via small manipulations of individual stock values that together realize an adversarial example to financial forecasting models, causing these models to make the self-fulfilling prediction of a crash. Such a crash triggered by an adversarial example would likely be hard to detect, since the model's predictions would be accurate and the interventions that would cause it are minor. This possibility is a major risk to financial stability and an opportunity for hostile actors to cause great economic damage to an adversary. This threat also exists against individual stocks and the corresponding valuation of individual companies. We outline how such an attack might proceed, what its theoretical basis is, how it can be directed towards a whole economy or an individual company, and how one might defend against it. We conclude that this threat is vastly underappreciated and requires urgent research on how to defend against it.

As large language models are increasingly deployed in sensitive environments, fingerprinting attacks pose significant privacy and security risks. We present a study of LLM fingerprinting from both offensive and defensive perspectives. Our attack methodology uses reinforcement learning to automatically optimize query selection, achieving better fingerprinting accuracy with only 3 queries compared to randomly selecting 3 queries from the same pool. Our defensive approach employs semantic-preserving output filtering through a secondary LLM to obfuscate model identity while maintaining semantic integrity. The defensive method reduces fingerprinting accuracy across tested models while preserving output quality. These contributions show the potential to improve fingerprinting tools capabilities while providing practical mitigation strategies against fingerprinting attacks.