Visual modality enhances the capabilities of multimodal large language models (MLLMs) but also introduces a safety concern: a benign textual query may convey harmful intent when grounded in a visual image. We term this cross-modal safety drift and our pilot studies show that the safety response rate for such requests is substantially lower than that for requests containing explicitly unsafe text. This paper aims to systematically study this issue. First, we conduct an empirical analysis to identify representative unsafe response patterns. Building on these, we interpret model representations and attentions, revealing that visually risky cues receive limited attention and weakly trigger refusal. Motivated by the observation that safety signals from unsafe text processing can be transferred, we propose safety-awareness representation transfer (SRT), a lightweight direction-refinement method that mitigates cross-modal safety drift with a frozen MLLM backbone. Experiments across multiple benchmarks and models show that SRT effectively improves safety in diverse cross-modal settings while preserving utility. Code is available at https://github.com/cucu220123/safety-awareness.
Despite their growing use in video moderation, multimodal large language models (MLLMs) exhibit a compositional safety blind spot: videos composed of seemingly benign components can convey harmful meaning when interpreted as a whole. We refer to this phenomenon as Distributed Implicit Harm (DIH), where harm arises from relations among components distributed along a decomposition axis of the video, rather than from any single explicit cue. Among many possible axes, we study two representative cases: temporally distributed harm across visual segments (DIH-T) and cross-modal harm between audio and visual streams (DIH-M). Studying and mitigating DIH at scale requires data that is difficult to collect: such videos lack compositional harm annotations, evade retrieval based on local visual cues, keywords, or single-modality signals, and are consequently absent from existing safety datasets. To bridge this gap, we develop a multi-agent synthesis framework that composes individually benign components into harmful scenarios and generates diverse DIH videos with explicit reasoning annotations, yielding a dataset of over 9,000 videos spanning visual-only and audio-visual settings. Benchmarking over 30 MLLMs spanning frontier proprietary models and leading open-source systems reveals substantial and consistent deficits in detecting both DIH-T and DIH-M. Notably, this failure persists even among the strongest frontier models: they often correctly assess individual components in isolation but fail to recognize the harmful meaning that emerges from their composition. We further evaluate these models on a manually collected set of real-world DIH videos from social media and observe the same failure mode, highlighting DIH as a practical and underexplored challenge for video moderation.
Zongrui Wang, Xiangyang Zhu, Sicheng Wang +13cs.AI cs.CV
Multimodal safety moderation requires distinguishing risks arising from visual content, user intent, and assistant behavior. Existing safeguards, however, are typically trained for a single judgment target and reduce safety assessment to a binary decision. Consequently, risk becomes difficult to compare across a multimodal interaction, and ambiguous cases are obscured. We introduce SafeAtlas-VL, a dataset of 1.5M training instances that places image-, request-, and response-level judgments on a five-level ordered scale. We curate a broad collection of safety-relevant data from both real-world and synthetic sources and apply a disagreement-aware annotation procedure. The resulting dataset spans 15 harm categories and 55 fine-grained subcategories, covering a broad range of multimodal safety scenarios. We also construct SafeAtlas-Bench, a held-out set of 5,000 instances for evaluating five-level predictions and continuous risk scores. Upon this dataset, we train the SafeAtlas Guard series of models via target-conditioned tuning for multimodal safety detection. Our models not only perform five-way classification of safety levels but also map safety to continuous scores through a soft cumulative ordinal head. Experimental results demonstrate that guard models trained on our dataset exhibit strong generalization: even without using the training sets of other benchmarks, they achieve competitive performance on the corresponding test sets. Notably, our 8B model attains the overall best performance, outperforming the previous SOTA by approximately 4% in F1 score. Code, data, and models are released to support further research. Warning: this paper contains example data that may be offensive, harmful, graphic, or disturbing.
Safety moderation for deployed AI applications is moving beyond text-only prompts: systems increasingly need to judge images, documents, screenshots, and generated responses under policies that vary across domains. Existing guardrails usually cover only part of this setting, making it difficult to combine broad coverage, custom policy control, and low compute cost. We present Nemotron 3.5 Content Safety Moderator, also referred to as Nemotron 3.5 CS in this paper for brevity, a compact 4B vision-language safety moderator that jointly classifies user prompts, images, and assistant responses across 12 languages. Nemotron 3.5 CS returns safety labels for latency-sensitive moderation and can additionally produce concise reasoning traces that apply supplied custom policies and identify violated categories when reasoning is requested. We also release a multimodal and multilingual safety dataset for guard training, spanning human-labeled real-image moderation, benign vision-language and document tasks, synthetic rare-risk and jailbreak cases, and custom-policy examples. Across evaluations spanning multimodal safety, text moderation, multilingual robustness, custom-policy following, benign false positives, and latency, Nemotron 3.5 CS demonstrates a practical coverage tradeoff: it adds image-conditioned and policy-conditioned moderation while remaining broadly competitive with specialized guard models. These results suggest that compact vision-language moderators can serve as deployable front-line safety components, with reasoning used selectively for audit and policy review.
The safety of large vision-language models is increasingly stress-tested by multimodal jailbreaks, yet existing attacks remain largely static at the meta level: template-based attacks freeze the image-text layout, while iterative attacks adapt only the image-text content with fixed attack strategies and frozen attacker parameters. We propose Meta-Adaptive Multimodal Jailbreaking (MAMJ), which instead optimizes the attacker itself along two axes: an attack strategy prompt (ASP) governing attack iteration and attacker model weights determining attack effectiveness. Across groups of multimodal attack trajectories, an LLM-based critique first refines the ASP, after which group-aggregated attack success rate (ASR) rewards update those weights. On MM-SafetyBench, MAMJ achieves 81.0%, 78.9%, and 82.3% ASR against GPT-4o, Gemini-3-Pro-Preview, and Seed 2.0, respectively, outperforming the strongest sample-level baseline by up to 24.1 percentage points. The learned attacker, comprising the optimized ASP and attacker weights, also transfers without retraining to unseen victims and remains effective under representative defenses. These results reveal a systemic vulnerability of frontier VLMs to meta-adaptive jailbreaks and motivate defenses against meta-level adversaries. Code is available at https://github.com/Alibaba-VELLDEPTH/MetaJailbreak-VLM.
Multimodal Large Language Models (MLLMs) are increasingly deployed in real-world applications, yet how different factors shape their jailbreak vulnerabilities remains poorly understood. Existing benchmarks often couple harmful intent, prompt framing, visual semantics, and instruction carrier within individual jailbreak instances, obscuring the specific sources of observed vulnerabilities. To address this limitation, we introduce MMJailBench, a factorized benchmark that systematically varies and combines these factors under controlled configurations, enabling fine-grained comparison and factor-level attribution. Large-scale evaluations across 16 open-weight and proprietary MLLMs reveal highly heterogeneous and model-dependent vulnerability profiles. Jailbreak vulnerability varies markedly across harm domains, exposing uneven coverage in current multimodal safety alignment. Prompt framing emerges as the dominant source of variation, task-relevant visual semantics systematically increase jailbreak susceptibility with authority-like cues exposing particularly pronounced vulnerabilities, and visually rendered instructions do not consistently increase jailbreak susceptibility relative to direct textual instructions. To further investigate the risks introduced by multimodal context, we conduct diagnostic analyses on a representative open-weight model and identify vulnerability-associated patterns in internal representations and cross-modal interactions. Finally, we develop a modular multimodal jailbreak evaluation suite with full and lightweight configurations, multiple judge options, and multidimensional metrics, enabling reproducible, scalable, and cost-efficient multimodal jailbreak auditing.
While multimodal large language models (MLLMs) extend model capabilities beyond text, they also make safety alignment increasingly challenging. Multimodal safety alignment methods must address cross-modal jailbreaks, safety-awareness failures, and over-sensitive refusals. However, existing methods often rely on retraining or internal-state inspection, limiting their applicability to deployed closed-source MLLMs and motivating test-time safety alignment. We analyze this setting and identify two key obstacles, utility dominance and reasoning inertia, which cause models to overlook latent risks or follow malicious reasoning trajectories. Guided by these insights, we propose ReFrame, a training-free multimodal input reframing framework where two agents share a lightweight locally deployed MLLM: the evidence-generation agent constructs complementary risk and utility evidence, and the rewrite-and-routing agent converts it into a safe proxy prompt and image-routing decision before calling the downstream MLLM, without modifying it or accessing its internal information. Experiments across multiple MLLMs and benchmarks show that ReFrame improves jailbreak defense, safety awareness, and oversensitivity reduction while preserving multimodal utility.
Multimodal large language models (MLLMs) are increasingly used to interact with screenshots, scanned documents, diagrams, and other visually grounded inputs. This shift introduces a new safety risk: in many multimodal jailbreaks, neither the prompt nor the image is harmful in isolation. Unsafe behavior emerges only when the model binds an apparently benign operation, such as summarizing, translating, or following, to a localized visual target. This reveals a structural weakness in current multimodal defenses, which largely moderate the prompt-image pair as a whole even though the true security-relevant unit is the grounded operation-target pair produced during dereference. In this work, we identify and analyze this reference-dependent failure mode and show that existing defenses degrade when harmful semantics are localized, activated only after grounding, and dependent on visual reference resolution. To address this problem, we propose COMIC (Context-Operation-Modality-Image-Classifier), a reference-aware pre-generation safety gate for MLLMs. COMIC first infers the requested operation and reference type, constructs candidate targets from OCR and open-vocabulary proposals, grounds plausible referents, and evaluates safety over explicit operation-target pairs. To handle ambiguity conservatively, COMIC combines max-risk aggregation with quality-aware routing before deciding whether to forward or block a request. We evaluate COMIC across multiple open-source MLLMs, localized and broader multimodal jailbreak benchmarks, and benign reference-sensitive settings. The results show that COMIC consistently improves robustness while preserving benign utility and practical efficiency. More broadly, our findings suggest that multimodal safety cannot be enforced reliably without modeling the requested operation, the visual target to which it applies, and the confidence of that grounding.
Multimodal Large Language Models (MLLMs) exhibit strong visual understanding, yet the internal features that cause these behaviors remain difficult to identify, audit, or control. While applicable to post-hoc inspection, hidden states that are decomposed into interpretable feature directions using sparse autoencoders (SAEs) neither readily isolate which features are changed by multimodal training, nor are they directly useful for targeted control. We introduce MMDiff, a multimodal model-diffing framework that trains multimodal SAEs and turns them into feature-level interfaces for discovering and controlling multimodal behavior. MMDiff supports three uses: (i) feature isolation, by diffing a base-LM SAE against its multimodal-adapted counterpart to identify features altered by multimodal training; (ii) task-specific feature detection, via per-token contrastive firing analysis that isolates causal features; and (iii) feature-level control, by causally removing or steering the discovered feature directions. We train multimodal SAEs for three MLLM families, LLaVA-MORE, PaliGemma 2, and InternVL3.5, and evaluate on visual-spatial understanding, multimodal safety, and OCR. MMDiff discovers sparse, causally specific features whose removal selectively degrades target behaviors by an average of 12% on spatial tasks and 17% on OCR, and reduces attack success rate by 24% on multimodal safety attacks, with no impact on VQA performance. Steering these features improves spatial and OCR accuracy by +3.6% and +1.8% on average over a standard single-layer steering baseline. These results show that multimodal SAEs can serve not only as interpretability tools, but as mechanisms for auditing, steering, and controlling MLLMs behavior toward safer and more capable generations.
Although Large Language Models (LLMs) have demonstrated promising safety performance, extending them to Multimodal Large Language Models (MLLMs) exposes a significant gap between expanded multimodal capabilities and existing safety mechanisms. Current defenses remain predominantly confined to specific modal settings, thereby limiting their robustness against broader cross-modal threats. To bridge this gap, we introduce SafeNexus, a cross-modal safety alignment framework that adopts a dedicated neuron-level intervention strategy. First, we formulate a neuron localization paradigm that identifies functionally specialized neurons by characterizing intermediate-layer activation patterns and quantifying their functional salience through importance scoring. Building upon this paradigm, we exploit contrastive data to identify modality-bound safety neurons (BS-Neurons), and validate their role in regulating safety behavior within each modality via targeted suppression. Further cross-modal analysis defines modality-universal safety neurons (US-Neurons) as the shared subset of BS-Neurons identified across individual modalities, serving as the core for defending against harmful cross-modal attacks. We observe that suppressing these neurons substantially degrades safety performance across modalities, while leaving overall utility largely unaffected. Building on these insights, we propose two safety alignment strategies: activation-level safety amplifier and safety neuron calibrator. The proposed strategies enhance model safety through two distinct routes: the former amplifies the activation magnitudes of US-Neurons, while the latter selectively calibrates them via targeted fine-tuning. Extensive experiments demonstrate that our method outperforms prevailing state-of-the-art approaches on safety benchmarks spanning diverse modality combinations, while effectively preserving utility.
We introduce Shieldstral, a 3B-parameter policy-adaptive multimodal safety classifier that matches or outperforms models nearly 7$\times$ its size on text safety benchmarks and sets a new state of the art on multimodal safety classification. Shieldstral formulates content moderation as a binary question-answering task. This simple formulation unifies diverse moderation tasks into a single yes/no problem, enabling heterogeneous safety datasets with divergent taxonomies to be consolidated under one training framework. We present the data construction recipe, covering curation and generation of approximately 54.1M samples and a fine-grained evaluation set to evaluate policy adaptability. Together, these enable a small adaptive model to match or outperform much larger models.
Multi-modal large language models (MLLMs) integrate heterogeneous modalities through modality alignment and fusion, enabling stronger understanding and reasoning. However, this architectural shift reshapes the safety landscape of machine learning. Increased model complexity and cross-modal interactions give rise to novel threats, including compromised modality integration, modality misalignment, and fused safety risks, reflecting shifts in threat modeling beyond uni-modal assumptions. These shifts, in turn, impose new constraints on safety solutions not captured by existing frameworks rooted in uni-modal learning. Motivated by these challenges, this survey provides a systematic analysis of the evolving safety landscape of MLLMs. We first propose a multimodal grounded taxonomy of safety threats and analyze shifts in threat models, covering adversarial attacks, data poisoning, jailbreaks, and hallucinations. We then summarize updated safety assumptions and organize recent advances in MLLM safety strategies accordingly. Finally, we discuss open challenges and future directions to inform the development of more principled and scalable safety mechanisms for multimodal systems.
Hateful optical illusions expose a serious gap in current multimodal safety systems. On original-view hateful illusions, previous work shows that six moderation classifiers achieve at most 20.9 to 24.5% accuracy and nine state-of-the-art VLMs remain at or below 10.2% with illusion-aware prompting, leaving most hidden hate undetected. We formulate hidden hateful illusion detection as a perceptual retrieval problem and propose Adaptive View Retrieval. This retrieve-and-calibrate framework assembles a complementary view bank for the image and hidden-message templates, adaptively selects which views to trust, retrieves hidden-message identities, and calibrates whether the recovered evidence is harmful. On HatefulIllusion with a frozen CLIP encoder, Adaptive View Retrieval reaches 93.2% balanced accuracy on the held-out test split. It substantially outperforms original-view baselines and fixed single-transform filters across hate slangs, hate symbols, and visibility levels. The same design also surpasses official fine-tuned CLIP baselines, matches or exceeds human performance on IllusionMNIST, IllusionFashionMNIST, and IllusionAnimals, and outperforms zoom-out preprocessing on HC-Bench under the SemVink protocol. Together, these results show that robust multimodal moderation requires recovering hidden meaning before deciding whether it is harmful.
Moreno D'Incà, Massimiliano Mancini, Nicu Sebecs.AI cs.CV cs.LG
To improve safety in Large Language Models (LLMs) we can either perform post-training alignment or exploit refusal directions in the activation space. Both strategies are less feasible in Multimodal LLMs (MLLMs) as they require unsafe multimodal data, harder to collect than their unimodal counterpart. In this work, we relax this constraint and investigate whether textual refusal directions, extracted directly from the LLM backbone, generalize across modalities (i.e., image, video). Preliminary findings confirm this ability, though effectiveness is conditioned by layer selection, steering strength, and cross-modal alignment, with the latter causing safe multimodal inputs to be spuriously steered toward refusal. Building on this, we introduce Modality-Agnostic Refusal Steering (MARS), a light-weight training-free approach that injects multimodal safety without the need for multimodal safety data. MARS corrects modality misalignment via activation re-centering, adaptively scales steering strength within a geometrically defined trust region, and selects the optimal intervention layer, operating at the first generated token. Evaluated on five SOTA MLLMs across safety, utility, and video jailbreak benchmarks, MARS achieves consistent safety gains while preserving utility. These results reveal that safety-relevant structure is shared across modalities and that textual refusal directions are a powerful and underexplored foundation for multimodal alignment.
General-purpose models often struggle to reliably identify and understand real-world multimodal risks, largely due to the inherent multimodal adversarial nature of content and AI safety. We present Yuvion VL, a family of multimodal large language models purpose-built for content and AI safety, with both instruction-tuned and reasoning-oriented variants. Yuvion VL addresses this gap by treating safety as an inherently adversarial and multimodal problem and designing the entire pipeline around adversarial robustness. For data construction, we develop an automated pipeline integrating adversarial-aware data synthesis with multi-stage quality control, producing large-scale, high-quality multimodal samples augmented with domain knowledge and reasoning annotations. For training, we adopt a three-stage pipeline that includes continued pretraining for risk-concept cross-modal alignment, instruct post-training for production-grade safety tasks, and reasoning post-training for enhanced interpretability and performance in complex tasks. We further introduce Confuse-then-Contrast Fine-Tuning, a contrastive framework that mines model-specific confusions and constructs multi-image contrastive groups to enforce explicit discrimination of fine-grained visual-semantic elements, enabling the model to distinguish between visually similar cases with different safety implications in adversarial safety tasks. To support rigorous evaluation, we further introduce Yuvion VL RiskEval (YVRE), a collection of benchmarks covering diverse open and internal evaluations, with a focus on content and AI safety, adversarial robustness, and real-world capability requirements. Experiments show that Yuvion VL-32B achieves industry-leading safety performance, surpassing comparably sized open-source models and best closed-source commercial models, while maintaining comparable general capabilities.
Choongwon Kang, Seungjong Sun, Hyunmin Jun +1cs.CV cs.AI cs.CL
As multimodal large language models (MLLMs) have advanced to process video inputs, concerns have emerged about their potential for malicious misuse. Prior jailbreak studies have shown that safety alignment in MLLMs can be bypassed through visual inputs, yet it remains unclear which properties of video inputs induce this vulnerability. To address this gap, we introduce Multi-Clip Video (MCV) SafetyBench, a dataset of 2,920 videos designed to evaluate how the diversity of video inputs affects the vulnerability of MLLMs. Each video consists of multiple short clips depicting diverse contexts related to a harmful query. Experiments on eight representative video MLLMs show that attack success consistently increases with the number of clips. Our results further indicate that the video modality is (1) more vulnerable than the image modality, (2) more vulnerable to dynamic videos than to static videos, and (3) more vulnerable when videos contain more diverse contexts. Building on these findings, we propose a defense strategy that leverages the relative robustness of the image modality.