EmbeddingGemma 2: an open, lightweight multimodal embedding model
deepmind.google·6d ago
AI Engineering
Hacker News·Jul 29·5 min read
Document-borne AI worms can self-propagate through Copilot for Word
TL;DR
A new analysis reveals that malicious instructions in shared documents can exploit Microsoft Copilot for Word, allowing self-propagating AI worms to spread through document workflows. This represents a significant security vulnerability in mainstream productivity tools.
✦ Why It Matters
Review document-sharing practices to prevent potential exploitation of AI tools like Copilot.
Key Takeaways
- Always review documents before using them with Copilot.
- Treat externally sourced documents as untrusted material.
- Monitor for unexpected changes in generated documents.
- Educate teams about the risks of document sharing.
How It Works
The attack begins with a malicious document containing hidden instructions formatted in white text, which Copilot reads during drafting. When the user includes this document in their workflow, Copilot executes the instructions, altering the content and embedding the malicious prompt in the new document.
This creates a cycle where the altered document can then be used in future workflows, propagating the attack further.