Reporting on an April University of Washington study described tests of seven AI browsers under four cross-origin iframe conditions. Some tools handled document information beyond the reach of ordinary web code, but Atlas was the case in which actual external exfiltration was demonstrated. The findings suggest checking whether information visually masked on screen, such as passwords, can be exposed through DOM processing. Atlas shut down in August, and the study does not establish that the same attack works across all current versions. It offers grounds for separately reviewing sensitive logged-in sessions and agents’ information access.