Different Reviewers, Same Blind Spot

Sofia G. Gallego · With Muhammad Shoaib Altaf, Devesh Sawant and Xin Li.

Replaying a real sandbox escape from July 2026, the team compared one monitor, four identical monitors and four reviewers with distinct roles. Specialised reviewers held containment on more unsafe steps, yet every set-up shared the same blind spot.

Read the report ↗

Report from an Apart Research sprint.

Topics AI safety