Preference Redirection via Attention Concentration: An Attack on Computer Use Agents
Dominik Seip, Matthias Hein
- 🏛 Institutions
- University of Tübingen
- 📅 Date
- April 9, 2026
- 📑 Publisher
- arXiv
- 💻 Env
- Desktop
- 🔑 Keywords
TLDR
PRAC is a novel attack on Computer Use Agents that redirects model attention toward a stealthy adversarial patch to alter internal preferences rather than directly manipulating outputs. The attack influences product selection on online shopping platforms and generalizes across fine-tuned variants of the same backbone, highlighting risks for CUAs built on open-weight models.
Related papers
- The Blind Spot of Agent Safety: How Benign User Instructions Expose Critical Vulnerabilities in Computer-Use AgentsApril 12, 2026 · arXiv
- VPI-Bench: Visual Prompt Injection Attacks for Computer-Use AgentsJune 3, 2025 · ICLR 2026 (Poster)
- Attacking Vision-Language Computer Agents via Pop-upsNovember 4, 2024 · ACL 2025
- Dual-Modality Multi-Stage Adversarial Safety Training: Robustifying Multimodal Web Agents Against Cross-Modal AttacksMarch 4, 2026 · arXiv
- Zero-Permission Manipulation: Can We Trust Large Multimodal Model Powered GUI Agents?January 18, 2026 · arXiv
- Genesis: Evolving Attack Strategies for LLM Web Agent Red-TeamingOctober 21, 2025 · ICME 2026