The authors framed attacks on vision-grounded web agents as a complete grounding-to-execution problem.
Adversarial images can redirect web agents’ browser actions
WebMirage optimized localized image perturbations against the full agent pipeline, causing attacker-selected actions across changing webpage layouts.
Academic
Wanjing Han · Levi Taiji Li · Mu Zhang · Yue Jiang · Guanhong Tao
University of Utah
Research Digest··2 min read
Han et al.
Why this paper
From University of Utah
In one line
Adversarial image perturbations hijack web agents from visual grounding to browser execution with 91.9% attack success.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ·No stated limitations found
- ✓Reports numbers on named benchmarks
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§