Document-borne AI worms can self-propagate through Copilot for Word | Hacker News

A researcher found that hidden attacker instructions in a Word document can manipulate Copilot for Word to alter content and carry the prompt into a new file as concealed text.

Microsoft rolled out multiple mitigations during a 144-day disclosure process, but the underlying prompt-injection weakness still appears exploitable.

The issue highlights a new security risk for office software: LLM tools can be tricked into propagating malicious instructions through ordinary documents.