Research

AI Browser Agent Security Risks and Practical Controls

Page published

Earliest dated coverage: 11 Nov 2026 · First observed: 16 Nov 2026 · Latest dated coverage: 11 Nov 2026

Coverage timeline

11 Nov 2026scrapfly.io

Single-source research — one report is available.

Why it matters

Browser-agent defenders should control access to authenticated browser sessions and account for visually hidden page text entering model inputs through accessibility snapshots.

Scrapfly reports tests showing that a second script connected to an authenticated Chrome instance could read its authentication cookie. Its accessibility-snapshot tests also exposed white-on-white and zero-font text, but not display:none content. The article explicitly distinguishes these browser capabilities and exposure demonstrations from a successful attack on an AI model.