One reason we trust it more than manual approval: in a study with 1,053 paid tes

Repeated permission prompts can lose their value if users stop inspecting them. In Anthropic's controlled coding study, 1,053 paid testers encountered a permission prompt whose text had been replaced with a dangerous command. Approval never executed that command. Testers rejected it 13.6% of the time, compared with auto mode blocking 89% of the same commands, according to the company's study account.

Human detection fell from about 17% early in a session to about 5% after at least 50 previous prompts. Auto mode instead evaluates proposed actions with a classifier. Anthropic presents this as a reason to trust auto mode over manual approval, though auto mode still allowed about 11% of the dangerous commands in this controlled test.