Skip to content

One Attack to Fool Them All: Highly Transferable Black-Box Adversarial Attacks on Frontier MLLMs

Sep 2026 · arXiv.org · Vol abs/2609.33833 · 0 citations · 102 references
Computer Science

TL;DR

This work proposes O-Attack, a highly transferable black-box attack framework that outperforms six state-of-the-art methods in black-box transferability, with consistent effectiveness across prompts and improved efficiency and imperceptibility.

Abstract

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.

View source

Similar papers

Sep 2026

Towards Transferable Black-Box Attack via Minimizing Maximum Model Discrepancy.

Adversarial attacks on black-box models, which operate without direct access to the target system, present a significant challenge due to the lack of a foundational theory for the transferability of adversarial examples. This paper introduces a paradigm shift in black-box adversarial attacks by minimizing model discrep...

An-Qi Zhao, Tong Chu, Ya-Hao Liu et al. · 0 citations
#artificial intelligence Preprint Sep 2026

What Makes Adversarial Examples Transfer Across Deepfake Detectors?

A controlled evaluation of adversarial transferability across 60 detectors spanning six backbones, two pretraining regimes, and five training-data configurations establishes source-model selection as central dimensions of credible transfer-based black-box robustness evaluation.

Rafael M. Mamede, Pedro C. Neto, A. F. Sequeira · 0 citations
Open access 2026

Unmasking Composite Adversaries: A New Threat Model for Federated Learning

Federated learning (FL) is increasingly adopted in distributed and privacy-preserving applications for decentralized intelligence, yet it remains highly vulnerable to adversarial attacks. Existing defenses remain fundamentally limited to single-attack scenarios, which overlook the real-world risk of coordinated, multiv...

M. Hasan, Yong Xiang, Md. Palash Uddin et al. · 0 citations
Preprint Sep 2026

CrACK: Adversarial Attacks on Cross-Model Consistency in Collaborative Vision Foundation Models

This work proposes CrACK (Cross-model Adversarial Consistency attack), an inference-time attack that exploits this interface without modifying any input pixel, model weight, or training data, and shows that it causes catastrophic degradation while every individual model continues to produce its unchanged standalone out...

Fei-Fei Liu, Jin-Tao Cheng, Chi-man Vong et al. · 0 citations
#artificial intelligence Preprint Oct 2026

Target-free Latent Safety Alignment

A target-free adversarial training framework that generates adversarial samples in an unsupervised manner that elicits diverse harmful behaviors in the target model, substantially mitigating behavioral narrowness and improving robustness to jailbreak attacks.

Luo-Yu Chen, Wei-Qi Wang, Chen-Han Zhang et al. · 0 citations

Related blog posts

MIT News · Artificial Intelligence Oct 28, 2026

Making machines that can listen

Professor Paris Smaragdis develops computers with superhuman hearing for applications like speech recognition, music understanding, and environmental monitoring.

Microsoft Research Blog Aug 11, 2026

Introducing CARE-X: Towards Clinically Useful Radiology VLMs with Auxiliary Supervision, Reward-Aligned Learning, and Tool-Augmented Measurement

Radiology AI is evolving beyond report generation. CARE-X explores a unified approach that combines flexible reasoning, calibrated predictions, and measurement-based tools for chest X-ray interpretation. The post Introducing CARE-X: Towards Clinically Useful Radiology VLMs with Auxiliary Supervision, Reward-Aligned Learning, and Tool-Augmented Measurement appeared first on Microsoft Research.

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.