DarkLLM: Learning Language-Driven Adversarial Attacks with Large Language Models

Ye Sun, Xin Wang, Jiaming Zhang, Yifeng Gao, Yixu Wang +5 more
5/15/2026
cs.CRcs.AIcs.CV

No official code found. Best implementation: sunye23/DarkLLM (1 star, MIT, last commit 2026-05-20).

Abstract

While vision and multimodal foundation models underpin critical tasks from perception to complex reasoning, they remain highly vulnerable to adversarial attacks. However, traditional adversarial attacks are typically limited to single, predefined objectives, tightly coupling each attack to a specific model or task, which restricts their scalability and flexibility in real-world scenarios. In this work, we present DarkLLM, a novel attack framework that trains an LLM to translate natural-language attack instructions into latent attack vectors, which are then decoded into visual adversarial perturbations. By leveraging natural-language instruction tuning, DarkLLM not only unifies targeted, untargeted, segmentation, and multi-model attacks within a single framework, but also achieves flexible and controllable adversarial generation, enabling each instruction to produce a perturbation that induces desired behaviors across heterogeneous models. Through extensive experiments across 4 tasks, 13 datasets, and 15 models, we demonstrate that DarkLLM with only 1B parameters can follow attacker instructions and generate highly effective attacks against CLIP, SAM, and frontier LLMs, revealing a systemic vulnerability in modern foundation models.

View on arXivView PDF

Code Implementations(1)

sunye23/DarkLLMCommunity65%
MIT

The implementation of DarkLLM: Learning Language-Driven Adversarial Attacks with Large Language Models

10Feb 2, 20265 months agoMIT
adversarial-attacksclipcross-modal-learninglarge-language-modelsmulti-modal-large-language-model+1 more

Similar Papers

Visual Adversarial Attack on Vision-Language Models for Autonomous Driving

88% match2024

As Firm As Their Foundations: Can open-sourced foundation models be used to create adversarial examples for downstream tasks?

88% match1 repo32024

InstructTA: Instruction-Tuned Targeted Attack for Large Vision-Language Models

88% match1 repo92023

Safeguarding Vision-Language Models Against Patched Visual Prompt Injectors

87% match2024

Forbid Your Attention: Fooling Multimodal Large Language Models by Selectively Removing Intrinsic Focus in Spectral Domain

87% match2026

VQAttack: Transferable Adversarial Attacks on Visual Question Answering via Pre-trained Models

87% match1 repo92024

Cite this paper

@article{sun2026darkllm,
  title  = {DarkLLM: Learning Language-Driven Adversarial Attacks with Large Language Models},
  author = {Ye Sun and Xin Wang and Jiaming Zhang and Yifeng Gao and Yixu Wang and Yifan Ding and Qixian Zhang and Henghui Ding and Xingjun Ma and Yu-Gang Jiang},
  year   = {2026},
  eprint = {2605.18868},
  archivePrefix = {arXiv},
  url    = {http://arxiv.org/abs/2605.18868v1}
}

Add a badge to your README

Shows whether official code is linked for this paper.

Code of Paper badge
[![Code of Paper](https://codeofpaper.com/badge/2605.18868.svg)](https://codeofpaper.com/paper/2605.18868)

Discussion