Claude Code 預設開啟自動模式：安全數據顯示人類僅能識別 13% 的危險指令
anthropic 將於 8 月 14 日預設啟用 claude code 的自動模式，此前數據顯示其攔截了 89% 的危險指令，遠超人類的 13.6%，標誌著開發者從編寫代碼轉向審查 ai 輸出。
anthropic will default claude code to auto mode on august 14, allowing the ai to perform coding tasks autonomously without manual approval for every step. testing with 1,053 paid users revealed that auto mode's classifier detected 89% of dangerous actions, significantly outperforming human reviewers who caught only 13.6%. an independent audit by trajectory labs confirmed that auto mode withstood 720 prompt injection attempts across claude 3.5 sonnet, claude 3 opus, and claude 3 haiku models without a single breach. while anthropic positions this as a safety upgrade that increases output by 25%, it fundamentally alters the developer role from writing code to supervising autonomous agents, introducing risks where teams may lose deep understanding of codebases built primarily by ai.