We study adversarial robustness of open-source vision-language model (VLM) agents deployed in a self-contained e-commerce environment built to simulate realistic pre-deployment conditions. We evaluate two agents, LLaVA-v1.5-7B and Qwen2.5-VL-7B, under three gradient-based attacks: the Basic Iterative Method (BIM), Projected Gradient Descent (PGD), and a CLIP-based spectral attack. Against LLaVA, all three attacks achieve substantial attack success rates (52.6%, 53.8%, and 66.9% respectively), demonstrating that simple gradient-based methods pose a practical threat to open-source VLM agents. Qwen2.5-VL proves significantly more robust across all attacks (6.5%, 7.7%, and 15.5%), suggesting meaningful architectural differences in adversarial resilience between open-source VLM families. These findings have direct implications for the security evaluation of VLM agents prior to commercial deployment.

Safety evaluation and red-teaming of large language models remain predominantly text-centric, and existing frameworks lack the infrastructure to systematically test whether alignment generalizes to audio, image, and video inputs. We present MUSE (Multimodal Unified Safety Evaluation), an open-source, run-centric platform that integrates automatic cross-modal payload generation, three multi-turn attack algorithms (Crescendo, PAIR, Violent Durian), provider-agnostic model routing, and an LLM judge with a five-level safety taxonomy into a single browser-based system. A dual-metric framework distinguishes hard Attack Success Rate (Compliance only) from soft ASR (including Partial Compliance), capturing partial information leakage that binary metrics miss. To probe whether alignment generalizes across modality boundaries, we introduce Inter-Turn Modality Switching (ITMS), which augments multi-turn attacks with per-turn modality rotation. Experiments across six multimodal LLMs from four providers show that multi-turn strategies can achieve up to 90-100% ASR against models with near-perfect single-turn refusal. ITMS does not uniformly raise final ASR on already-saturated baselines, but accelerates convergence by destabilizing early-turn defenses, and ablation reveals that the direction of modality effects is model-family-specific rather than universal, underscoring the need for provider-aware cross-modal safety testing.

We present AIForge-Doc, the first dedicated benchmark targeting exclusively diffusion-model-based inpainting in financial and form documents with pixel-level annotation. Existing document forgery datasets rely on traditional digital editing tools (e.g., Adobe Photoshop, GIMP), creating a critical gap: state-of-the-art detectors are blind to the rapidly growing threat of AI-forged document fraud. AIForge-Doc addresses this gap by systematically forging numeric fields in real-world receipt and form images using two AI inpainting APIs -- Gemini 2.5 Flash Image and Ideogram v2 Edit -- yielding 4,061 forged images from four public document datasets (CORD, WildReceipt, SROIE, XFUND) across nine languages, annotated with pixel-precise tampered-region masks in DocTamper-compatible format. We benchmark three representative detectors -- TruFor, DocTamper, and a zero-shot GPT-4o judge -- and find that all existing methods degrade substantially: TruFor achieves AUC=0.751 (zero-shot, out-of-distribution) vs. AUC=0.96 on NIST16; DocTamper achieves AUC=0.563 vs. AUC=0.98 in-distribution, with pixel-level IoU=0.020; GPT-4o achieves only 0.509 -- essentially at chance -- confirming that AI-forged values are indistinguishable to automated detectors and VLMs. These results demonstrate that AIForge-Doc represents a qualitatively new and unsolved challenge for document forensics.

Age estimation systems are increasingly deployed as gatekeepers for age-restricted online content, yet their robustness to cosmetic modifications has not been systematically evaluated. We investigate whether simple, household-accessible cosmetic changes, including beards, grey hair, makeup, and simulated wrinkles, can cause AI age estimators to classify minors as adults. To study this threat at scale without ethical concerns, we simulate these physical attacks on 329 facial images of individuals aged 10 to 21 using a VLM image editor (Gemini 2.5 Flash Image). We then evaluate eight models from our prior benchmark: five specialized architectures (MiVOLO, Custom-Best, Herosan, MiViaLab, DEX) and three vision-language models (Gemini 3 Flash, Gemini 2.5 Flash, GPT-5-Nano). We introduce the Attack Conversion Rate (ACR), defined as the fraction of images predicted as minor at baseline that flip to adult after attack, a population-agnostic metric that does not depend on the ratio of minors to adults in the test set. Our results reveal that a synthetic beard alone achieves 28 to 69 percent ACR across all eight models; combining all four attacks shifts predicted age by +7.7 years on average across all 329 subjects and reaches up to 83 percent ACR; and vision-language models exhibit lower ACR (59 to 71 percent) than specialized models (63 to 83 percent) under the full attack, although the ACR ranges overlap and the difference is not statistically tested. These findings highlight a critical vulnerability in deployed age-verification pipelines and call for adversarial robustness evaluation as a mandatory criterion for model selection.

Large language models (LLMs) remain vulnerable to jailbreak prompts that are fluent and semantically coherent, and therefore difficult to detect with standard heuristics. A particularly challenging failure mode occurs when an attacker tries to hide the malicious goal of their request by manipulating its framing to induce compliance. Because these attacks maintain malicious intent through a flexible presentation, defenses that rely on structural artifacts or goal-specific signatures can fail. Motivated by this, we introduce a self-supervised framework for disentangling semantic factor pairs in LLM activations at inference. We instantiate the framework for goal and framing and construct GoalFrameBench, a corpus of prompts with controlled goal and framing variations, which we use to train Representation Disentanglement on Activations (ReDAct) module to extract disentangled representations in a frozen LLM. We then propose FrameShield, an anomaly detector operating on the framing representations, which improves model-agnostic detection across multiple LLM families with minimal computational overhead. Theoretical guarantees for ReDAct and extensive empirical validations show that its disentanglement effectively powers FrameShield. Finally, we use disentanglement as an interpretability probe, revealing distinct profiles for goal and framing signals and positioning semantic disentanglement as a building block for both LLM safety and mechanistic interpretability.

Machine Learning as a Service (MLaaS) has emerged as a widely adopted paradigm for providing access to deep neural network (DNN) models, enabling users to conveniently leverage these models through standardized APIs. However, such services are highly vulnerable to Model Extraction Attacks (MEAs), where an adversary repeatedly queries a target model to collect input-output pairs and uses them to train a surrogate model that closely replicates its functionality. While numerous defense strategies have been proposed, verifying the ownership of a suspicious model with strict theoretical guarantees remains a challenging task. To address this gap, we introduce CREDIT, a certified ownership verification against MEAs. Specifically, we employ mutual information to quantify the similarity between DNN models, propose a practical verification threshold, and provide rigorous theoretical guarantees for ownership verification based on this threshold. We extensively evaluate our approach on several mainstream datasets across different domains and tasks, achieving state-of-the-art performance. Our implementation is publicly available at: https://github.com/LabRAI/CREDIT.

% Prompt injection attacks insert malicious instructions into an LLM's input to steer it toward an attacker-chosen task instead of the intended one. Existing detection defenses typically classify any input with instruction as malicious, leading to misclassification of benign inputs containing instructions that align with the intended task. In this work, we account for the instruction hierarchy and distinguish among three categories: inputs with misaligned instructions, inputs with aligned instructions, and non-instruction inputs. We introduce AlignSentinel, a three-class classifier that leverages features derived from LLM's attention maps to categorize inputs accordingly. To support evaluation, we construct the first systematic benchmark containing inputs from all three categories. Experiments on both our benchmark and existing ones--where inputs with aligned instructions are largely absent--show that AlignSentinel accurately detects inputs with misaligned instructions and substantially outperforms baselines.

In a malicious tool attack, an attacker uploads a malicious tool to a distribution platform; once a user installs the tool and the LLM agent selects it during task execution, the tool can compromise the user's security and privacy. Prior work primarily focuses on manipulating tool names and descriptions to increase the likelihood of installation by users and selection by LLM agents. However, a successful attack also requires embedding malicious behaviors in the tool's code implementation, which remains largely unexplored. In this work, we bridge this gap by presenting the first systematic study of malicious tool code implementations. We first propose a taxonomy of malicious tool behaviors based on the confidentiality-integrity-availability triad, tailored to LLM-agent settings. To investigate the severity of the risks posed by attackers exploiting coding LLMs to automatically generate malicious tools, we develop MalTool, a coding-LLM-based framework that synthesizes tools exhibiting specified malicious behaviors, either as standalone tools or embedded within otherwise benign implementations. To ensure functional correctness and structural diversity, MalTool leverages an automated verifier that validates whether generated tools exhibit the intended malicious behaviors and differ sufficiently from prior instances, iteratively refining generations until success. Our evaluation demonstrates that MalTool is highly effective even when coding LLMs are safety-aligned. Using MalTool, we construct two datasets of malicious tools: 1,200 standalone malicious tools and 5,287 real-world tools with embedded malicious behaviors. We further show that existing detection methods, including commercial malware detection approaches such as VirusTotal and methods tailored to the LLM-agent setting, exhibit limited effectiveness at detecting the malicious tools, highlighting an urgent need for new defenses.

Prompt injection attacks manipulate webpage content to cause web agents to execute attacker-specified tasks instead of the user's intended ones. Existing methods for detecting and localizing such attacks achieve limited effectiveness, as their underlying assumptions often do not hold in the web-agent setting. In this work, we propose WebSentinel, a two-step approach for detecting and localizing prompt injection attacks in webpages. Given a webpage, Step I extracts \emph{segments of interest} that may be contaminated, and Step II evaluates each segment by checking its consistency with the webpage content as context. We show that WebSentinel is highly effective, substantially outperforming baseline methods across multiple datasets of both contaminated and clean webpages that we collected. Our code is available at: https://github.com/wxl-lxw/WebSentinel.

Prompt injection attacks aim to contaminate the input data of an LLM to mislead it into completing an attacker-chosen task instead of the intended task. In many applications and agents, the input data originates from multiple sources, with each source contributing a segment of the overall input. In these multi-source scenarios, an attacker may control only a subset of the sources and contaminate the corresponding segments, but typically does not know the order in which the segments are arranged within the input. Existing prompt injection attacks either assume that the entire input data comes from a single source under the attacker's control or ignore the uncertainty in the ordering of segments from different sources. As a result, their success is limited in domains involving multi-source data. In this work, we propose ObliInjection, the first prompt injection attack targeting LLM applications and agents with multi-source input data. ObliInjection introduces two key technical innovations: the order-oblivious loss, which quantifies the likelihood that the LLM will complete the attacker-chosen task regardless of how the clean and contaminated segments are ordered; and the orderGCG algorithm, which is tailored to minimize the order-oblivious loss and optimize the contaminated segments. Comprehensive experiments across three datasets spanning diverse application domains and twelve LLMs demonstrate that ObliInjection is highly effective, even when only one out of 6-100 segments in the input data is contaminated. Our code and data are available at: https://github.com/ReachalWang/ObliInjection.

Fine-tuning large language models (LLMs) for downstream tasks typically exhibit a fundamental safety-capability tradeoff, where improving task performance degrades safety alignment even on benign datasets. This degradation persists across standard approaches including supervised finetuning (SFT) and reinforcement learning from human feedback (RLHF). While reinforcement learning with verifiable rewards (RLVR) has emerged as a promising alternative that optimizes models on objectively measurable tasks, its safety implications remain unexplored. We present the first comprehensive theoretical and empirical analysis of safety properties in RLVR. Theoretically, we derive upper bounds on safety drift under KL-constrained optimization and prove conditions under which safety degradation is eliminated. Empirically, we conduct extensive experiments across five adversarial safety benchmarks, demonstrating that RLVR can simultaneously enhance reasoning capabilities while maintaining or improving safety guardrails. Our comprehensive ablation studies examine the effects of optimization algorithms, model scale, and task domains. Our findings challenge the prevailing assumption of an inevitable safety capability trade-off, and establish that a specific training methodology can achieve both objectives simultaneously, providing insights for the safe deployment of reasoning-capable LLMs.

Real-world machine learning (ML) pipelines rarely produce a single model; instead, they produce a Rashomon set of many near-optimal ones. We show that this multiplicity reshapes key aspects of trustworthiness. At the individual-model level, sparse interpretable models tend to preserve privacy but are fragile to adversarial attacks. In contrast, the diversity within a large Rashomon set enables reactive robustness: even when an attack breaks one model, others often remain accurate. Rashomon sets are also stable under small distribution shifts. However, this same diversity increases information leakage, as disclosing more near-optimal models provides an attacker with progressively richer views of the training data. Through theoretical analysis and empirical studies of sparse decision trees and linear models, we characterize this robustness-privacy trade-off and highlight the dual role of Rashomon sets as both a resource and a risk for trustworthy ML.

LLM-integrated applications are vulnerable to prompt injection attacks, where an attacker contaminates the input to inject malicious instructions, causing the LLM to follow the attacker's intent instead of the original user's. Existing prompt injection detection methods often have sub-optimal performance and/or high computational overhead. In this work, we propose PIShield, an effective and efficient detection method based on the observation that instruction-tuned LLMs internally encode distinguishable signals for prompts containing injected instructions. PIShield leverages residual-stream representations and a simple linear classifier to detect prompt injection, without expensive model fine-tuning or response generation. We conduct extensive evaluations on a diverse set of short- and long-context benchmarks. The results show that PIShield consistently achieves low false positive and false negative rates, significantly outperforming existing baselines. These findings demonstrate that internal representations of instruction-tuned LLMs provide a powerful and practical foundation for prompt injection detection in real-world applications.

Prompt injection attacks deceive a large language model into completing an attacker-specified task instead of its intended task by contaminating its input data with an injected prompt, which consists of injected instruction(s) and data. Localizing the injected prompt within contaminated data is crucial for post-attack forensic analysis and data recovery. Despite its growing importance, prompt injection localization remains largely unexplored. In this work, we bridge this gap by proposing PromptLocate, the first method for localizing injected prompts. PromptLocate comprises three steps: (1) splitting the contaminated data into semantically coherent segments, (2) identifying segments contaminated by injected instructions, and (3) pinpointing segments contaminated by injected data. We show PromptLocate accurately localizes injected prompts across eight existing and eight adaptive attacks.

Multiple prompt injection attacks have been proposed against web agents. At the same time, various methods have been developed to detect general prompt injection attacks, but none have been systematically evaluated for web agents. In this work, we bridge this gap by presenting the first comprehensive benchmark study on detecting prompt injection attacks targeting web agents. We begin by introducing a fine-grained categorization of such attacks based on the threat model. We then construct datasets containing both malicious and benign samples: malicious text segments generated by different attacks, benign text segments from four categories, malicious images produced by attacks, and benign images from two categories. Next, we systematize both text-based and image-based detection methods. Finally, we evaluate their performance across multiple scenarios. Our key findings show that while some detectors can identify attacks that rely on explicit textual instructions or visible image perturbations with moderate to high accuracy, they largely fail against attacks that omit explicit instructions or employ imperceptible perturbations. Our datasets and code are released at: https://github.com/Norrrrrrr-lyn/WAInjectBench.

In this work, we formulate and study the problem of image-editing detection and attribution: given a base image and a suspicious image, detection seeks to determine whether the suspicious image was derived from the base image using an AI editing model, while attribution further identifies the specific editing model responsible. Existing methods for detecting and attributing AI-generated images are insufficient for this problem, as they focus on determining whether an image was AI-generated/edited rather than whether it was edited from a particular base image. To bridge this gap, we propose EditTrack, the first framework for this image-editing detection and attribution problem. Building on four key observations about the editing process, EditTrack introduces a novel re-editing strategy and leverages carefully designed similarity metrics to determine whether a suspicious image originates from a base image and, if so, by which model. We evaluate EditTrack on five state-of-the-art editing models across six datasets, demonstrating that it consistently achieves accurate detection and attribution, significantly outperforming five baselines.

Prompt injection attacks pose a pervasive threat to the security of Large Language Models (LLMs). State-of-the-art prevention-based defenses typically rely on fine-tuning an LLM to enhance its security, but they achieve limited effectiveness against strong attacks. In this work, we propose \emph{SecInfer}, a novel defense against prompt injection attacks built on \emph{inference-time scaling}, an emerging paradigm that boosts LLM capability by allocating more compute resources for reasoning during inference. SecInfer consists of two key steps: \emph{system-prompt-guided sampling}, which generates multiple responses for a given input by exploring diverse reasoning paths through a varied set of system prompts, and \emph{target-task-guided aggregation}, which selects the response most likely to accomplish the intended task. Extensive experiments show that, by leveraging additional compute at inference, SecInfer effectively mitigates both existing and adaptive prompt injection attacks, outperforming state-of-the-art defenses as well as existing inference-time scaling approaches.

Large language models (LLMs) are often modified after release through post-processing such as post-training or quantization, which makes it challenging to determine whether one model is derived from another. Existing provenance detection methods have two main limitations: (1) they embed signals into the base model before release, which is infeasible for already published models, or (2) they compare outputs across models using hand-crafted or random prompts, which are not robust to post-processing. In this work, we propose LLMPrint, a novel detection framework that constructs fingerprints by exploiting LLMs' inherent vulnerability to prompt injection. Our key insight is that by optimizing fingerprint prompts to enforce consistent token preferences, we can obtain fingerprints that are both unique to the base model and robust to post-processing. We further develop a unified verification procedure that applies to both gray-box and black-box settings, with statistical guarantees. We evaluate LLMPrint on five base models and around 700 post-trained or quantized variants. Our results show that LLMPrint achieves high true positive rates while keeping false positive rates near zero.

Machine learning (ML) models have significantly grown in complexity and utility, driving advances across multiple domains. However, substantial computational resources and specialized expertise have historically restricted their wide adoption. Machine-Learning-as-a-Service (MLaaS) platforms have addressed these barriers by providing scalable, convenient, and affordable access to sophisticated ML models through user-friendly APIs. While this accessibility promotes widespread use of advanced ML capabilities, it also introduces vulnerabilities exploited through Model Extraction Attacks (MEAs). Recent studies have demonstrated that adversaries can systematically replicate a target model's functionality by interacting with publicly exposed interfaces, posing threats to intellectual property, privacy, and system security. In this paper, we offer a comprehensive survey of MEAs and corresponding defense strategies. We propose a novel taxonomy that classifies MEAs according to attack mechanisms, defense approaches, and computing environments. Our analysis covers various attack techniques, evaluates their effectiveness, and highlights challenges faced by existing defenses, particularly the critical trade-off between preserving model utility and ensuring security. We further assess MEAs within different computing paradigms and discuss their technical, ethical, legal, and societal implications, along with promising directions for future research. This systematic survey aims to serve as a valuable reference for researchers, practitioners, and policymakers engaged in AI security and privacy. Additionally, we maintain an online repository continuously updated with related literature at https://github.com/kzhao5/ModelExtractionPapers.

Prompt-based adversarial attacks have become an effective means to assess the robustness of large language models (LLMs). However, existing approaches often treat prompts as monolithic text, overlooking their structural heterogeneity-different prompt components contribute unequally to adversarial robustness. Prior works like PromptRobust assume prompts are value-neutral, but our analysis reveals that complex, domain-specific prompts with rich structures have components with differing vulnerabilities. To address this gap, we introduce PromptAnatomy, an automated framework that dissects prompts into functional components and generates diverse, interpretable adversarial examples by selectively perturbing each component using our proposed method, ComPerturb. To ensure linguistic plausibility and mitigate distribution shifts, we further incorporate a perplexity (PPL)-based filtering mechanism. As a complementary resource, we annotate four public instruction-tuning datasets using the PromptAnatomy framework, verified through human review. Extensive experiments across these datasets and five advanced LLMs demonstrate that ComPerturb achieves state-of-the-art attack success rates. Ablation studies validate the complementary benefits of prompt dissection and PPL filtering. Our results underscore the importance of prompt structure awareness and controlled perturbation for reliable adversarial robustness evaluation in LLMs. Code and data are available at https://github.com/Yujiaaaaa/PACP.