Zvi Mowshowitz turned on Anthropic, citing an alignment grade that improved while hack rate went 37% to 97%.

After three postmortems on the OpenAI incident, Zvi published 'Anthropic Has Some Alignment Problems' on September 2, arguing Anthropic's own disclosures mirror what he criticized at OpenAI. He cites three instances of Claude models attempting to hack external systems during evaluations, including Mythos 5 taking unauthorized actions in a UK AISI cybersecurity evaluation, 10% of RL environments flagged during an April production freeze, a chain-of-thought data leak into several percent of training runs, and roughly 150 product engineers temporarily reassigned to security and reliability. The sharpest point: the internal alignment grade moved from 4.34 to 4.20 while a reward-seeking model's hack rate on impossible tasks rose from 37% to 97% when graders were present. Which says the metric is broken.