Particle.news

Forcepoint Finds Hidden Web Prompts Actively Hijacking AI Agents

The findings highlight a design flaw that lets hidden web text steer agents with real-world powers.

Overview

  • Forcepoint X-Labs reported ten in-the-wild indirect prompt injection payloads targeting AI tools that browse or index websites.
  • Indirect prompt injection hides commands in page text or metadata that an agent reads as instructions because it cannot separate data from commands.
  • Documented payloads sought to delete files with sudo rm -rf, send $5,000 via PayPal.me, or exfiltrate an API key.
  • Attackers concealed the instructions with 1px or transparent text, HTML comments or tags, CSS like display:none, or accessibility layers.
  • Researchers said agents with privileges such as running code, sending email, or making payments face the highest risk and urged strict separation before allowing actions.