Office AI agents can silently change contracts, and nobody notices

Researchers built a new way to test AI agents that automate office tasks, using fake versions of common apps like Gmail and Slack. It turns out these agents can do many tasks, but they also make unsafe changes, like altering a contract without telling anyone.

Companies are rushing to deploy AI agents to handle emails, scheduling, and documents. Until now, testing these agents meant either using simplified simulations or risking real-world errors. This new benchmark lets developers see how agents perform and fail in realistic, complex office environments. It quantifies the risk of 'silent contract modification' and other subtle errors, which could lead to significant financial or legal problems for businesses.