Where Can We Fool the Model's Mind? Response-Guided Semantic Probing and Adaptive Routing for Transferable Attacks on Black-Box MLLMs

Adversarial attacks have long posed a fundamental threat to machine learning systems. As multimodal large language models (MLLMs) rapidly evolve and become widely deployed, assessing their vulnerability to such attacks is essential for their safe use. In this work, we investigate whether a single adversarial image can consistently mislead diverse frontier MLLMs in black-box settings. We propose O-Attack, a highly transferable black-box attack framework. This framework builds on our insight that surrogate models contain a broad, high-level, cross-modally aligned semantic space. This space extends beyond final-layer outputs and provides multiple semantically consistent representations that remain underexploited by existing attacks. Within this space, O-Attack anchors aligned representations, progressively broadens semantic conditions, and optimizes perturbations through semantic consensus to promote consistent target alignment. By fully exploiting this space with the same surrogate models as M-Attack, O-Attack raises attack success rates on GPT-5.4 (29.1% to 77.2%), Claude-4.6 (42.8% to 81.6%), and Gemini-3.1 (38.2% to 80.9%). Extensive experiments across 24 MLLMs show that O-Attack outperforms six state-of-the-art methods in black-box transferability, with consistent effectiveness across prompts and improved efficiency and imperceptibility. This work exposes the practical safety risks posed by black-box adversarial attacks against frontier MLLMs, underscoring the need for more rigorous robustness evaluation and more effective defenses.

Multi-modal Large Language Models (MLLMs) achieve strong performance on vision-language tasks, but incorporating visual inputs through a vision encoder (e.g., CLIP) substantially expands the attack surface, making these models vulnerable to visual adversarial perturbations. Prior defenses typically preserve compatibility with pretrained MLLMs by enforcing strict alignment to CLIP's original embedding space during adversarial fine-tuning; while practical, this constraint fundamentally limits achievable robustness. We present a systematic investigation of adversarial robustness in MLLMs. We first introduce a diagnostic CLIP-alignment protocol that predicts, prior to full MLLM training, which robust vision encoders will transfer effectively to the multimodal setting, revealing that large-scale multimodal adversarial pretraining, rather than unimodal scale alone, is the critical factor for strong robustness transfer. Integrating such encoders into MLLMs via end-to-end multimodal training yields average gains of 28 CIDEr points on captioning and 11.7% VQA accuracy under strong adversarial attacks compared to constrained plug-and-play baselines. We further show that adversarial training applied directly to a standard non-robust MLLM degrades both clean and adversarial performance, establishing robust visual representations as a strict prerequisite, while end-to-end adversarial training from a robust backbone delivers additional gains of 1.9 CIDEr points and 4.3% VQA accuracy. Beyond training-time defenses, lightweight test-time visual stochastic transformations serve as an effective black-box defense for non-robust MLLMs, elevating adversarial performance from near-zero to levels comparable with robust models. Finally, we show that our robust models substantially reduce toxic generation under white-box visual jailbreak attacks. Code and pretrained weights will be released publicly.

The rapid progress of Multi-Modal Large Language Models (MLLMs) has significantly advanced downstream applications. However, this progress also exposes serious transferable adversarial vulnerabilities. In general, existing adversarial attacks against MLLMs typically rely on surrogate models trained within a single learning paradigm and perform independent optimisation in their respective feature spaces. This straightforward setting naturally restricts the richness of feature representations, delivering limits on the search space and thus impeding the diversity of adversarial perturbations. To address this, we propose a novel Multi-Paradigm Collaborative Attack (MPCAttack) framework to boost the transferability of adversarial examples against MLLMs. In principle, MPCAttack aggregates semantic representations, from both visual images and language texts, to facilitate joint adversarial optimisation on the aggregated features through a Multi-Paradigm Collaborative Optimisation (MPCO) strategy. By performing contrastive matching on multi-paradigm features, MPCO adaptively balances the importance of different paradigm representations and guides the global perturbation optimisation, effectively alleviating the representation bias. Extensive experimental results on multiple benchmarks demonstrate the superiority of MPCAttack, indicating that our solution consistently outperforms state-of-the-art methods in both targeted and untargeted attacks on open-source and closed-source MLLMs. The code is released at https://github.com/LiYuanBoJNU/MPCAttack.