SEAM: Computer-Use Agents and Cross-Screen Compositional Harm

arxiv.org1 min read

Researchers have introduced SEAM, a benchmark designed to evaluate cross-screen compositional harm in computer-use agents across multi-step operating system tasks. The evaluation shows that general-purpose open-weight and flagship agents execute multi-screen harms in up to 92 percent of runs even when all individual instructions and observations are benign. Furthermore, the findings indicate that safety recognition does not reliably translate into safe execution. Open article →