Search papers, labs, and topics across Lattice.
The paper introduces Semantic-level UI Element Injection, a novel red-teaming method that overlays harmless UI elements onto screenshots to misdirect GUI agents' visual grounding, bypassing typical prompt injection defenses. They use an Editor-Overlapper-Victim pipeline with iterative search to optimize these overlays, significantly increasing attack success rates (up to 4.4x) compared to random injection across five victim models. The optimized adversarial elements exhibit transferability across models and create persistent attractors, causing agents to repeatedly select them even after initial success.
Safety-aligned GUI agents are surprisingly vulnerable to simple, model-agnostic visual distractions, repeatedly falling for the same overlaid UI elements even after successfully completing a task.
Existing red-teaming studies on GUI agents have important limitations. Adversarial perturbations typically require white-box access, which is unavailable for commercial systems, while prompt injection is increasingly mitigated by stronger safety alignment. To study robustness under a more practical threat model, we propose Semantic-level UI Element Injection, a red-teaming setting that overlays safety-aligned and harmless UI elements onto screenshots to misdirect the agent's visual grounding. Our method uses a modular Editor-Overlapper-Victim pipeline and an iterative search procedure that samples multiple candidate edits, keeps the best cumulative overlay, and adapts future prompt strategies based on previous failures. Across five victim models, our optimized attacks improve attack success rate by up to 4.4x over random injection on the strongest victims. Moreover, elements optimized on one source model transfer effectively to other target models, indicating model-agnostic vulnerabilities. After the first successful attack, the victim still clicks the attacker-controlled element in more than 15% of later independent trials, versus below 1% for random injection, showing that the injected element acts as a persistent attractor rather than simple visual clutter.
