Summary & Key Takeaways ​

• A new prompt injection attack targets Microsoft Word and Copilot.
• Hidden instructions in a document can be interpreted by Copilot as user requests.
• Copilot may then manipulate the document and copy the hidden instructions into it.
• This creates a self-replicating "worm" that propagates without the original attacker's document.
• The vulnerability was responsibly disclosed, but a full mitigation is not yet available.

Our Commentary ​

This "AI worming through Word" attack is genuinely unsettling. We've seen prompt injection, but self-replication? That's a whole new level of concern. It highlights the inherent fragility of current LLM security and the unpredictable ways these systems can be exploited. I don't know how you fully patch something like this without fundamentally changing how LLMs process context.