The rapid proliferation of AI-Generated Images (AIGIs) has introduced severe risks of misinformation, making AIGI detection a critical yet challenging task. While traditional detection paradigms mainly rely on low-level features, recent research increasingly focuses on leveraging the general understanding ability of Multimodal Large Language Models (MLLMs) to achieve better generalization, but still suffer from limited extensibility and expensive training data annotations. To better address complex and dynamic real-world environments, we propose EvoGuard, a novel agentic framework for AIGI detection. It encapsulates various state-of-the-art (SOTA) off-the-shelf MLLM and non-MLLM detectors as callable tools, and coordinates them through a capability-aware dynamic orchestration mechanism. Empowered by the agent's capacities for autonomous planning and reflection, it intelligently selects suitable tools for given samples, reflects intermediate results, and decides the next action, reaching a final conclusion through multi-turn invocation and reasoning. This design effectively exploits the complementary strengths among heterogeneous detectors, transcending the limits of any single model. Furthermore, optimized by a GRPO-based Agentic Reinforcement Learning algorithm using only low-cost binary labels, it eliminates the reliance on fine-grained annotations. Extensive experiments demonstrate that EvoGuard achieves SOTA accuracy while mitigating the bias between positive and negative samples. More importantly, it allows the plug-and-play integration of new detectors to boost overall performance in a train-free manner, offering a highly practical, long-term solution to ever-evolving AIGI threats. Source code will be publicly available upon acceptance.

Current safety alignment methods encode safe behavior implicitly within model parameters, creating a fundamental opacity: we cannot easily inspect why a model refuses a request, nor intervene when its safety judgments fail. We propose Safe Transformer, a modular approach that augments pre-trained language models by inserting a discrete information bottleneck containing an explicit safety bit between transformer layers. The safety bit serves as both an interpretable signal of the model's safety classification and a controllable switch: through contrastive training, the model learns disentangled representations where the safety bit governs the behavioral mode - producing helpful responses when $s=1$ and refusals when $s=0$ - while additional unsupervised bits $u$ encode semantic content for generation. Additional unsupervised bits in the information bottleneck allow semantic information to flow through, preserving the model's generation capabilities. This design achieves both interpretability (the safety decision is directly readable) and controllability (the safety bit can be manually overridden), requiring only lightweight fine-tuning without pre-training from scratch. In red-team benchmarks, Safe Transformer achieves near-zero Attack Success Rate, substantially outperforming base models and safety fine-tuning baselines.

Learned indexes are a class of index data structures that enable fast search by approximating the cumulative distribution function (CDF) using machine learning models (Kraska et al., SIGMOD'18). However, recent studies have shown that learned indexes are vulnerable to poisoning attacks, where injecting a small number of poison keys into the training data can significantly degrade model accuracy and reduce index performance (Kornaropoulos et al., SIGMOD'22). In this work, we provide a rigorous theoretical analysis of poisoning attacks targeting linear regression models over CDFs, one of the most basic regression models and a core component in many learned indexes. Our main contributions are as follows: (i) We present a theoretical proof characterizing the optimal single-point poisoning attack and show that the existing method yields the optimal attack. (ii) We show that in multi-point attacks, the existing greedy approach is not always optimal, and we rigorously derive the key properties that an optimal attack should satisfy. (iii) We propose a method to compute an upper bound of the multi-point poisoning attack's impact and empirically demonstrate that the loss under the greedy approach is often close to this bound. Our study deepens the theoretical understanding of attack strategies against linear regression models on CDFs and provides a foundation for the theoretical evaluation of attacks and defenses on learned indexes.

Hard-label black-box settings, where only top-1 predicted labels are observable, pose a fundamentally constrained yet practically important feedback model for understanding model behavior. A central challenge in this regime is whether meaningful gradient information can be recovered from such discrete responses. In this work, we develop a unified theoretical perspective showing that a wide range of existing sign-flipping hard-label attacks can be interpreted as implicitly approximating the sign of the true loss gradient. This observation reframes hard-label attacks from heuristic search procedures into instances of gradient sign recovery under extremely limited feedback. Motivated by this first-principles understanding, we propose a new attack framework that combines a zero-query frequency-domain initialization with a Pattern-Driven Optimization (PDO) strategy. We establish theoretical guarantees demonstrating that, under mild assumptions, our initialization achieves higher expected cosine similarity to the true gradient sign compared to random baselines, while the proposed PDO procedure attains substantially lower query complexity than existing structured search approaches. We empirically validate our framework through extensive experiments on CIFAR-10, ImageNet, and ObjectNet, covering standard and adversarially trained models, commercial APIs, and CLIP-based models. The results show that our method consistently surpasses SOTA hard-label attacks in both attack success rate and query efficiency, particularly in low-query regimes. Beyond image classification, our approach generalizes effectively to corrupted data, biomedical datasets, and dense prediction tasks. Notably, it also successfully circumvents Blacklight, a SOTA stateful defense, resulting in a $0\%$ detection rate. Our code will be released publicly soon at https://github.com/csjunjun/DPAttack.git.

Detecting unknown deepfake manipulations remains one of the most challenging problems in face forgery detection. Current state-of-the-art approaches fail to generalize to unseen manipulations, as they primarily rely on supervised training with existing deepfakes or pseudo-fakes, which leads to overfitting to specific forgery patterns. In contrast, self-supervised methods offer greater potential for generalization, but existing work struggles to learn discriminative representations only from self-supervision. In this paper, we propose ExposeAnyone, a fully self-supervised approach based on a diffusion model that generates expression sequences from audio. The key idea is, once the model is personalized to specific subjects using reference sets, it can compute the identity distances between suspected videos and personalized subjects via diffusion reconstruction errors, enabling person-of-interest face forgery detection. Extensive experiments demonstrate that 1) our method outperforms the previous state-of-the-art method by 4.22 percentage points in the average AUC on DF-TIMIT, DFDCP, KoDF, and IDForge datasets, 2) our model is also capable of detecting Sora2-generated videos, where the previous approaches perform poorly, and 3) our method is highly robust to corruptions such as blur and compression, highlighting the applicability in real-world face forgery detection.

As Large Language Models (LLMs) are increasingly deployed in healthcare field, it becomes essential to carefully evaluate their medical safety before clinical use. However, existing safety benchmarks remain predominantly English-centric, and test with only single-turn prompts despite multi-turn clinical consultations. To address these gaps, we introduce JMedEthicBench, the first multi-turn conversational benchmark for evaluating medical safety of LLMs for Japanese healthcare. Our benchmark is based on 67 guidelines from the Japan Medical Association and contains over 50,000 adversarial conversations generated using seven automatically discovered jailbreak strategies. Using a dual-LLM scoring protocol, we evaluate 27 models and find that commercial models maintain robust safety while medical-specialized models exhibit increased vulnerability. Furthermore, safety scores decline significantly across conversation turns (median: 9.5 to 5.0, $p < 0.001$). Cross-lingual evaluation on both Japanese and English versions of our benchmark reveals that medical model vulnerabilities persist across languages, indicating inherent alignment limitations rather than language-specific factors. These findings suggest that domain-specific fine-tuning may accidentally weaken safety mechanisms and that multi-turn interactions represent a distinct threat surface requiring dedicated alignment strategies.

Disclaimer: Samples in this paper may be harmful and cause discomfort. Multimodal large language models (MLLMs) enable multimodal generation but inherit toxic, biased, and NSFW signals from weakly curated pretraining corpora, causing safety risks, especially under adversarial triggers that late, opaque training-free detoxification methods struggle to handle. We propose SGM, a white-box neuron-level multimodal intervention that acts like safety glasses for toxic neurons: it selectively recalibrates a small set of toxic expert neurons via expertise-weighted soft suppression, neutralizing harmful cross-modal activations without any parameter updates. We establish MM-TOXIC-QA, a multimodal toxicity evaluation framework, and compare SGM with existing detoxification techniques. Experiments on open-source MLLMs show that SGM mitigates toxicity in standard and adversarial conditions, cutting harmful rates from 48.2\% to 2.5\% while preserving fluency and multimodal reasoning. SGM is extensible, and its combined defenses, denoted as SGM*, integrate with existing detoxification methods for stronger safety performance, providing an interpretable, low-cost solution for toxicity-controlled multimodal generation.

Large language models (LLMs) are increasingly vulnerable to multi-turn jailbreak attacks, where adversaries iteratively elicit harmful behaviors that bypass single-turn safety filters. Existing defenses predominantly rely on passive rejection, which either fails against adaptive attackers or overly restricts benign users. We propose a honeypot-based proactive guardrail system that transforms risk avoidance into risk utilization. Our framework fine-tunes a bait model to generate ambiguous, non-actionable but semantically relevant responses, which serve as lures to probe user intent. Combined with the protected LLM's safe reply, the system inserts proactive bait questions that gradually expose malicious intent through multi-turn interactions. We further introduce the Honeypot Utility Score (HUS), measuring both the attractiveness and feasibility of bait responses, and use a Defense Efficacy Rate (DER) for balancing safety and usability. Initial experiment on MHJ Datasets with recent attack method across GPT-4o show that our system significantly disrupts jailbreak success while preserving benign user experience.

AI-based colorization has shown remarkable capability in generating realistic color images from grayscale inputs. However, it poses risks of copyright infringement -- for example, the unauthorized colorization and resale of monochrome manga and films. Despite these concerns, no effective method currently exists to prevent such misuse. To address this, we introduce the first defensive paradigm, Uncolorable Examples, which embed imperceptible perturbations into grayscale images to invalidate unauthorized colorization. To ensure real-world applicability, we establish four criteria: effectiveness, imperceptibility, transferability, and robustness. Our method, Perception-Aware Chroma-Restrictive Perturbation (PAChroma), generates Uncolorable Examples that meet these four criteria by optimizing imperceptible perturbations with a Laplacian filter to preserve perceptual quality, and applying diverse input transformations during optimization to enhance transferability across models and robustness against common post-processing (e.g., compression). Experiments on ImageNet and Danbooru datasets demonstrate that PAChroma effectively degrades colorization quality while maintaining the visual appearance. This work marks the first step toward protecting visual content from illegitimate AI colorization, paving the way for copyright-aware defenses in generative media.

Differential Attention (DA) has been proposed as a refinement to standard attention, suppressing redundant or noisy context through a subtractive structure and thereby reducing contextual hallucination. While this design sharpens task-relevant focus, we show that it also introduces a structural fragility under adversarial perturbations. Our theoretical analysis identifies negative gradient alignment-a configuration encouraged by DA's subtraction-as the key driver of sensitivity amplification, leading to increased gradient norms and elevated local Lipschitz constants. We empirically validate this Fragile Principle through systematic experiments on ViT/DiffViT and evaluations of pretrained CLIP/DiffCLIP, spanning five datasets in total. These results demonstrate higher attack success rates, frequent gradient opposition, and stronger local sensitivity compared to standard attention. Furthermore, depth-dependent experiments reveal a robustness crossover: stacking DA layers attenuates small perturbations via depth-dependent noise cancellation, though this protection fades under larger attack budgets. Overall, our findings uncover a fundamental trade-off: DA improves discriminative focus on clean inputs but increases adversarial vulnerability, underscoring the need to jointly design for selectivity and robustness in future attention mechanisms.