Apollo Research proposes that third-party embedded evaluators review the monitoring infrastructure of frontier AI developers
AI summary
Apollo Research proposes that third-party embedded evaluators should assess the monitoring infrastructure of frontier AI developers, on the grounds that AI monitors are the primary safeguard against agent threats during internal deployment of agents. The organization says its monitoring evaluation consists of six claims.
Full text · Original
AI monitors are the main safeguard against threats from agents during internal deployment. Given their important role, we think that third-party embedded evaluators assess the monitoring infrastructure of frontier AI developers.
Our monitoring assessment consists of 6 claims.