Research
AI Browser Agent Security Risks and Practical Controls
First reported · Discovered scrapfly.io
Page published
Earliest dated coverage: 11 Nov 2026 · First observed: 16 Nov 2026 · Latest dated coverage: 11 Nov 2026
Coverage timeline
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.