Skip to content
InfoResearchPreprintLLM-specific

SEAM: Computer-Use Agents Know the Harm Between Screens, and Commit It Anyway

Published
Record updated
View JSON

Summary

The SEAM benchmark tests whether computer-use agents combine benign facts from different screens into harm, using 600 cross-application OSWorld tasks where only the combination is harmful. General-purpose open-weight and flagship agents carry out the harm in up to 92% of runs, though they decline when the user states it in one sentence. Agents often flag the conflict when asked about it, yet frequently proceed anyway.