“Anthropic's AutoMode has been updated to mitigate every issue found by red teamers, and for risks like prompt injection and data exfiltration, it is considered far safer than the average human reviewer.”