Håkon Måløy discovered a novel prompt injection technique that transforms malicious instructions in Microsoft Word documents into self-replicating AI worms via Copilot for Word. The attack propagates automatically when infected documents are reused in Copilot workflows, creating a persistent threat without requiring the original attacker's file.
Background
This vulnerability exploits how Copilot for Word processes hidden instructions within documents, turning standard prompt injection into an autonomous propagation mechanism. It represents a significant evolution in AI-powered document attacks beyond traditional data exfiltration methods.
- Source
- Simon Willison
- Published
- Jul 30, 2026 at 02:43 AM
- Score
- 8.0 / 10