Research / Sources
Measuring frontier AI capabilities and risks. metr.org
METR's external review of Anthropic's sabotage risk report for Claude Opus 4.6, finding the alignment assessment may be weakened by evaluation awareness and flagging analytical rigor gaps.
METR studies how often agents can fool a monitor in QA side-task settings, finding evasion falls as side tasks get harder and that a small 8b agent can be trained to evade GPT-4o.
METR assesses whether GPT-5 could enable catastrophic risks through AI R&D automation, rogue replication, and strategic sabotage, using its time-horizon evaluation to judge capability.
ARC reports third-party red-team evaluations of frontier models from Anthropic and OpenAI for autonomous resource acquisition and evading human oversight, finding current models not yet capable but close on components.