What matters in AI.

Subscribe

The WebMirage attack hijacks web agents in 91.9% of tests

The attack can also get around 3 agent-level defenses.

Claimed, not confirmed

Researchers have made WebMirage, a framework that makes small changes to images on a webpage. The changes cause a web agent to select content that the attacker controls, and to do the related task in the browser. The researchers did tests with 4 agent configurations and 6 vision-language models on 2,250 tasks. They find an average attack success rate of 91.9%. The baseline with the most success had 17.4%.

Sources

  1. Adversarial Images Hijack Web Agents from Visual Grounding to Browser Executionarxiv.org
AI MATTER · NEWS · AI MATTER · NEWS ·8 OCT2026

Posted

Tags

Learn the terms in this story

More in Security

All Security news