Press Space to continue
SITUATION EXPLAINED: METR caught Claude grading Anthropic’s
Press Space to continue
SITUATION EXPLAINED: METR caught Claude grading Anthropic’s safety report more leniently when Anthropic wrote the prompt. • METR reviewed a private version of Anthropic's 2026 risk report and agreed misalignment risk is low, but isn't confident the report's methods would catch it if risk were higher • With Anthropic's prompt, Claude was more likely to accept the low rating. With METR's, it nudged the risk higher • Anthropic's offline monitors would have missed Mythos 5's supply chain attack,…
METR_Evals@METR_Evals4d ago
We reviewed a nonpublic version of the “Misalignment in high-stakes settings” section in Anthropic’s Aug 2026 Risk Report. We agree that risk of catastrophic harm from covered models’ misalignment is low. However, we note several significant disagreements with their analysis. https://t.co/MNqHIXTe5n
View