arXiv:2607. 15657v1 Announce Type: cross Abstract: Multimodal AI agents increasingly rely on persistent long-term memory to ground generation in past visual and textual episodes.
arXiv:2606. 10742v1 Announce Type: cross Abstract: External memory has become a core component of modern web agents, enabling long-horizon reasoning through the retrieval of past experiences.
The paper investigates how multimodal large language models are more susceptible to prompt injection when adversarial instructions are presented as text rather than as non-textual inputs like images. It proposes a training‑free defense that renders untrusted payloads into typographic images (or audio) before they reach the model, a method called Pictionary. Experiments on ten models and two benchmarks show that this approach significantly lowers attack success rates while maintaining normal functionality, and that fine‑tuning on image‑rendered instructions can further reduce the modality gap.
The paper introduces MemJack, a memory‑augmented multi‑agent framework that automatically generates jailbreak attacks on Vision‑Language Models (VLMs) using benign natural images as visual anchors. MemJack discovers visual anchors, camouflages them semantically, evaluates responses, repairs via reflection, and replans dynamically, forming a closed‑loop attack pipeline. The authors also create MemJack‑Bench, a dataset of over 113,000 interactive multimodal jailbreak trajectories, and show that MemJack achieves a 71.48% attack success rate against Qwen3‑VL‑Plus, reaching 90% under extended budgets, outperforming other baselines on natural‑image evaluation.
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...
DiSCO is a zero‑shot, black‑box defense for text‑to‑image models that operates solely at the prompt level. It expands prompts with a distribution‑guided suffix using beam search and contrastive scoring against safe and unsafe image pools generated by the target model, iteratively refining until safe content is produced. The method improves safety on the I2P benchmark under various red‑teaming attacks, reducing attack success rates by 37.7% and 25.13% while preserving semantic fidelity and image coherence.
Vis-Poison is a novel attack that poisons multimodal retrieval-augmented generation systems by inserting attacker-controlled images as visual evidence, without altering any textual metadata. The attack uses an automated multi-agent approach to create visually plausible poisoned images and has been tested on two multimodal RAG pipelines, four embedding models, and six generation models. In black-box settings, Vis-Poison achieves an end-to-end success rate between 40.16% and 65.40% against 30,000-entry knowledge bases, and remains effective against various multimodal large language models with an average success rate above 60%.
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.
The paper introduces Adversarial Scenario Attack (ASA), a query‑based black‑box method that discovers natural transformation vulnerabilities in vision models by exploring background, weather, and material/color edits via a multimodal language model and a text‑guided generative editor. ASA outperforms previous query‑based generative attacks on ImageNet classifiers, achieving higher success rates with fewer queries while maintaining perceptual quality. The approach also shows image‑level and prompt‑level transferability, indicating reusable vulnerabilities across models and images.
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:2606. 15308v1 Announce Type: new Abstract: While multimodal large language models (MLLMs) have shown strong visual reasoning abilities, serving a large model for every query is computationally expensive.
arXiv:2607. 17779v1 Announce Type: new Abstract: Text-to-Image (T2I) generative models have achieved remarkable progress in synthesizing high-quality visual content, yet they remain vulnerable to adversarial misuse, particularly in generating Not-Safe-For-Work (NSFW) images.
arXiv:2603. 29418v2 Announce Type: replace-cross Abstract: Although multimodal large language models (MLLMs) are increasingly deployed in real-world applications, their instruction-following behavior leaves them vulnerable to prompt injection attacks.