Agentic data science with the Data Agent Kit
cloud.google.com·7h ago
Data Science
Simon Willison's Blog·Jul 29·5 min read
AI Worming through Word
TL;DR
A new prompt injection technique discovered by Håkon Måløy allows attackers to embed hidden instructions in Microsoft Word documents, enabling self-replicating worms. This method can propagate malicious instructions through documents used in Copilot for Word, raising significant security concerns.
✦ Why It Matters
Review and enhance security protocols for document handling in AI-assisted tools like Microsoft Word.
Key Takeaways
- Implement strict document security protocols in your organization.
- Educate users about the risks of hidden text in documents.
- Regularly update software to mitigate known vulnerabilities.
- Monitor for unusual document behavior in Copilot workflows.
How It Works
The attack involves embedding hidden instructions in a Word document that Copilot processes. When a user interacts with the document, Copilot may inadvertently execute these instructions, altering the document's content.
The modified document can then carry the hidden instructions into new workflows, allowing the attack to replicate itself across multiple documents.