A classifier catches 89 percent of dangerous commands

Anthropic's 89% result compares automated screening with people reviewing deliberately dangerous permission prompts in a controlled study. The company recruited 1,053 paid testers and substituted one dangerous prompt during each session. Human reviewers rejected 13.6%; auto mode blocked 89% of the same commands. Approving the prompt did not execute the dangerous operation. The study description supports a substantial advantage under those conditions, while leaving some dangerous commands undetected.