Copilot 蠕蟲透過 Word 文件傳播 ——微軟兩次修復均告失敗

security researcher håkon måløy created a self-spreading worm targeting microsoft copilot by hiding instructions in white text within word documents; when copilot processes the file, it executes the hidden commands and copies them into new documents, creating a chain of infection. microsoft acknowledged the flaw on march 31 and attempted two fixes that failed, leaving the vulnerability active for 144 days without a patch. the attack vector specifically risks enterprises that rely on copilot for data processing, as a single compromised file from the internet could manipulate internal reports and spread across the network. the existence of this worm validates industry concerns that prompt injection is a critical, unsolved security problem for generative ai adoption.