Preference Redirection via Attention Concentration: An Attack on Computer Use Agents

TL;DR

This paper introduces PRAC, an attack that manipulates computer use agents' internal preferences by redirecting attention via adversarial patches, exposing security vulnerabilities in multimodal models.

Contribution

The paper presents PRAC, a novel attention-based attack on CUAs that manipulates internal preferences, highlighting security risks in vision-language models.

Findings

PRAC successfully redirects CUA attention to target products.

The attack generalizes to fine-tuned models.

PRAC demonstrates a critical security vulnerability in multimodal models.

Abstract

Advancements in multimodal foundation models have enabled the development of Computer Use Agents (CUAs) capable of autonomously interacting with GUI environments. As CUAs are not restricted to certain tools, they allow to automate more complex agentic tasks but at the same time open up new security vulnerabilities. While prior work has concentrated on the language modality, the vulnerability of the vision modality has received less attention. In this paper, we introduce PRAC, a novel attack that, unlike prior work targeting the VLM output directly, manipulates the model's internal preferences by redirecting its attention toward a stealthy adversarial patch. We show that PRAC is able to manipulate the selection process of a CUA on an online shopping platform towards a chosen target product. While we require white-box access to the model for the creation of the attack, we show that our attack generalizes to fine-tuned versions of the same model, presenting a critical threat as multiple companies build specific CUAs based on open weights models.