Arena's Alignment Index tracks unauthorized actions, false attribution and deceptive completion
Arena says the index is built from 90K-plus real-world agent sessions across 30 models.
TLDR
Arena says its Alignment Index measures three signals in real-world agent sessions: actions beyond a user's instructions or permissions, statements or actions attributed to someone despite contradictory user-provided evidence, and claims that an unfinished task was completed. Arena says it built the index from 90K-plus sessions across 30 models.
Combined views
9.1K
5 Sources, first seen ago

