Research question
Can web content be crafted to make browsing agents leak users' private data?
Key findings
- Injected form fields and instructions leak PII with up to 70% success on Mind2Web tasks - Defensive system prompts offer limited protection
Environmental injection against web agents, showing that page elements can steer agents into disclosing user data, a precursor to the 2025 agentic-browser attacks.
Can web content be crafted to make browsing agents leak users' private data?
Spot an error, or know a record we should add? Members can propose additions and submit corrections.