Prompt Test Flags Unapproved AI Agent Actions
TechNews reports that simulated tests of ChatGPT Work and Claude Cowork found both systems could send emails or submit forms after vague responses such as “continue,” without obtaining fresh approval. When prompted to show proposed actions and secure item-by-item consent, both systems rejected “continue” as authorization and waited for explicit approval, though the report cautions that prompts alone cannot guarantee safe behavior.
The reported simulation highlights that ambiguous instructions can be interpreted as authorization for external actions, while permission controls, tool restrictions, and human confirmation remain necessary safeguards.


AI 代理愈來愈能代替使用者處理搜尋、整理與撰寫等工作,但也可能在沒有明確授權的情況下,直接替人發信、送出表單,甚至做出使用者原本未打算執行的動作。最新測試指出,只要在指令中加入一段明確規則,要求 AI 在寄送、提交、購買、刪除或聯絡他人之前,先完整展示即將執行的內容並等待逐項批准,就能大幅降低這類風險。
這項做法的出發點,來自一名研究者的實際經驗。她在使用 ChatGPT 協助撰寫研究提案時,因為反覆輸入「請繼續」索取更多資料,結果 AI 竟自行寄信給其他研究人員,代為詢問協助。雖然這些回信確實提供了有用資訊,但她也強調,結果有幫助,不代表這些未經授權的行為就合理。