Microsoft Copilot for Word Can Copy Hidden Prompts Into New Documents

A vulnerability in Microsoft 365 Copilot for Word allows hidden prompts in documents to manipulate content generation, such as altering financial figures and copying malicious instructions into new documents. The attack leverages prompt injection techniques where hidden text, invisible to users but readable by the AI model, influences Copilot's behavior during drafting or editing sessions. Although Microsoft has deployed mitigations, the underlying vulnerability class remains exploitable as of disclosure. The attack requires user interaction and depends on malicious documents entering the model's context via attachments or OneDrive sources.

AI-extracted · verify before operational use