What's this about?
Researchers disclosed (coordinated with Microsoft's MSRC) a vulnerability class where hidden instructions inside a shared Word document can hijack Copilot: it edits your documents as the attacker wants and propagates the malicious instructions into new documents — a self-spreading AI worm. At publication there was no robust mitigation for the broader class.
What exactly is a "prompt-injection worm"?
Prompt injection is when text that an AI reads as *data* (a document, a webpage) secretly contains *instructions* the AI then obeys. Make those instructions say "copy yourself into the next document you touch," and you get a worm: it spreads document-to-document through an AI that can't tell content from commands.
What's the community saying?
HN 350 pts, 268 comments
- "Why can Word even have hidden text the AI reads but the human can't see?" — calls to strip invisible content before it reaches Copilot.
- Alarm that there's no robust fix for the class yet, only case-by-case patches.
- Debate over whether this is truly the first AI worm or has priors.