Research
MIRAGE: Stealthy Visual Prompt Injection for Vulnerability Detection in Web Agents
Publication date unknown · Discovered arxiv.org
Page published
Publication date unknown · First observed: 31 Oct 2026
Coverage timeline
Single-source research — one report is available.
Why it matters
MIRAGE demonstrates why screenshot-driven web agents must treat third-party visual content as untrusted even when it appears within an otherwise trusted website.
MIRAGE is a visual indirect prompt-injection framework that uses adversarial images confined to third-party webpage regions, such as ad slots or widgets, to hijack web agents’ next actions. The paper reports evaluations against SeeAct and OpenClaw, using diffusion-generated images combined with sparse dark-pixel perturbations to make the attacks visually inconspicuous.