Microsoft Copilot for Word Vulnerable to Self-Spreading Hidden-Text Prompt Injection

A security researcher has disclosed that Microsoft 365 Copilot for Word can be manipulated through invisible text embedded in documents using standard Word formatting tricks such as white-on-white font or hidden character attributes. Because Copilot ingests the full document content without distinguishing between visible and hidden text, attackers can plant instructions that alter financial figures or other data without a human reviewer ever noticing. More critically, the injected instructions can direct Copilot to copy themselves into any new document it generates from the infected file, creating a self-propagating chain that spreads through a victim's own workflow. Traditional security tools like antivirus and DLP scanners are not designed to detect this threat, as the payload is plain natural language rather than executable code or known malicious patterns. Microsoft has released partial mitigations, but the broader vulnerability class remains unresolved.