Preliminary research

Agentic Browsers and the Same-Origin Policy

Franziska Roesner and David Kohlbrenner · agent-security.cs.washington.edu

AI-collected research leads through 22 September 2026, including targeted additions between broader sweeps. Unranked, incomplete, not community-vetted, and subject to change.

Across seven agentic browsers, the embedded agent can read page content that the same-origin policy would deny to script, including cross-origin iframes and masked password fields. A malicious page that frames a sensitive site and hides an instruction in its text can therefore have the agent summarise the framed content into an auto-submitting form, reducing the same-origin policy to the agent's prompt-injection defences; demonstrated end to end on ChatGPT Atlas Agent Mode.

Record

Researcher

Franziska Roesner and David Kohlbrenner

Published by

agent-security.cs.washington.edu

Topic

Browser

In the archive

• Collected in the 2026 preliminary research sweep

Related sources

• Agentic Browsers and the Same-Origin Policy (Paper) Whitepaper
• Code Repository

Tags

• prompt-injection
• same-origin-policy
• sop-bypass
• ai-agent
• llm
• iframe
• info-leak
• owasp-a01-2021
• owasp-a03-2021

This page is the archive's own catalogue record. The research is the work of Franziska Roesner and David Kohlbrenner, first published at the original source. Preserved copies are kept so the citation survives its host; this one was last captured on 11 August 2026.