The original title is about hidden prompt injection attacks on browser agents and testing defenses. Let me rewrite this for a mobile feed.

Short summary

A hands-on security lab demonstrating indirect prompt injection attacks against browser agents, where untrusted web page content can trick an agent into performing unauthorized actions. The exercise walks through building a harmless attacking page, recording agent actions, and then applying trust, capability, and approval controls to harden the agent. Results are compared before and after defensive restrictions using the user's own runs rather than claimed benchmarks.

• •Browser agents can be tricked by hostile web page content into performing unauthorized state-changing actions via indirect prompt injection
• •Lab provides reproducible setup with attacking page, audit log, vulnerable and hardened policy profiles
• •Defenses include treating page content as untrusted, limiting destinations/methods, and requiring approval for consequential actions

Generated with AI, which can make mistakes.