The paper introduces Multi-Modal Typography, a systematic investigation into how perturbations applied to speech and on-screen text can compromise audio‑visual reasoning in large language models. Experiments on the WorldSense benchmark show that such attacks can raise the attack success rate against Gemini 3.7 Flash from 9.53% to 43.41%, and fine‑tuning can reduce the success rate for Qwen2.5‑Omni‑7B from 60.6% to 52.0%. The study highlights that cross‑modal typography attacks are a significant, yet underexplored, vulnerability in multi‑modal reasoning systems.
arXiv:2609.20850v1 Announce Type: new Abstract: While Multimodal Large Language Models (MLLMs) show remarkable advancements, their cross-modal capabilities introduce complex vulnerabilities that easi...
arXiv:2603.02482v2 Announce Type: replace Abstract: Safety evaluation of multimodal large language models requires tracking not only whether an attack succeeds, but also how the interaction unfolds a...
arXiv:2608. 07535v1 Announce Type: cross Abstract: Multi-modal large language models (MLLMs) integrate heterogeneous modalities through modality alignment and fusion, enabling stronger understanding and reasoning.
The paper introduces OverThink, a slowdown attack that forces reasoning language models (RLMs) to produce many more reasoning tokens while still giving correct answers. By injecting decoy reasoning problems—such as Markov decision processes, language translation, or graphic comprehension—into the model’s context, attackers can dramatically increase token generation (up to 46× on SQuAD and 17× on coding agents). The study evaluates the attack on both proprietary and open-source RLMs across multiple datasets, explores multimodal and coding‑agent variants, and tests several defenses, concluding that defending against OverThink is challenging and that newer RLMs are even more vulnerable due to higher per‑token costs and increased reasoning token usage.
The paper investigates how Vision Language Models (VLMs) can be fooled by tiny, human‑imperceptible changes to images. It introduces a gradient‑based attack that targets only the vision encoder, reducing computational cost while still effectively disrupting both untargeted and targeted multimodal interpretations. Experiments on open‑source VLMs such as Qwen2.5‑VL, Granite‑Vision, FastVLM, and Phi‑3.5‑Vision demonstrate that these small perturbations can dramatically alter the models’ textual outputs.
arXiv:2511.18921v2 Announce Type: replace Abstract: Backdoor attacks undermine the reliability and trustworthiness of machine learning systems by injecting hidden behaviors that can be maliciously ac...
The paper investigates how Vision Language Models (VLMs) can be fooled by small, human‑imperceptible changes to images. It introduces a gradient‑based attack that targets only the vision encoder, reducing computational cost while still effectively disrupting both untargeted and targeted multimodal alignment. Experiments on open‑source VLMs such as Qwen2.5‑VL, Granite‑Vision, FastVLM, and Phi‑3.5‑Vision demonstrate that these perturbations can significantly alter the models’ textual outputs.
FusionAudit introduces PaRMA, a pathway-conditioned robustness auditing framework for native multimodal vision‑language models. It evaluates two image‑level interventions—pixel perturbation and rendered‑text insertion—to expose architecture‑specific blind spots. The study applies diagnostic‑motivated attacks (MANTIS and Nat‑Evid) to 18 open‑weight VLMs, revealing that legacy single‑attack rankings are unstable and that PaRMA uncovers higher residual risk than standard benchmarks.
ReFrame is a training‑free framework that enhances safety alignment for multimodal large language models at test time. It uses two lightweight agents: one generates risk and utility evidence, and the other rewrites prompts and routes images to create a safe proxy before invoking the deployed MLLM. Experiments show that ReFrame improves jailbreak defense, safety awareness, and reduces over‑sensitivity while maintaining multimodal utility.
Omni‑SafetyBench is a new benchmark designed to evaluate the safety of Omni‑Modal Large Language Models (OLLMs) that process visual, auditory, and textual data. It contains 23,328 test instances across 24 modality variations derived from 972 seed samples, and introduces metrics such as Safety‑score (based on Conditional Attack Success Rate and Conditional Refusal Rate) and Cross‑Modal Safety Consistency score. Evaluation of 11 state‑of‑the‑art OLLMs shows severe vulnerabilities, with only three models achieving a Safety‑score above 0.6 and safety degrading sharply for audio‑visual inputs, underscoring the need for improved safety alignment methods.
The paper introduces COMIC, a reference‑aware safety gate designed for multimodal large language models (MLLMs). COMIC detects the operation requested by a user, identifies visual targets through OCR and open‑vocabulary proposals, and evaluates safety on explicit operation‑target pairs, using max‑risk aggregation and quality‑aware routing to decide whether to allow or block a request. Experiments on several open‑source MLLMs and jailbreak benchmarks show that COMIC improves robustness while maintaining benign utility and efficiency.
MobileInject is a benchmark that evaluates the security of multimodal large language model (MLLM)-based mobile agents against multimodal injection attacks. It offers a high‑fidelity interactive framework with 11 scenarios, 9 risk categories, and 108 executable risk instances, and introduces Cross‑Modal Consistency Injection along with two advanced attack methods—Multi‑Role Collaborative Deception and Reasoning‑Camouflage Injection—to intensify trust manipulation and reasoning hijacking. Experiments show that mobile agents are broadly vulnerable, especially to early‑stage reasoning hijacking, establishing MobileInject as a systematic tool for studying and improving mobile agent security.