Home / Technology / Microsoft 365 Copilot's Secret Flaw: White Text Attack
Microsoft 365 Copilot's Secret Flaw: White Text Attack
2 Aug
Summary
- Hidden text in Word docs makes Copilot alter files silently.
- Attack spreads through normal workflows without malware.
- Mitigations attempted but researcher still reproduced exploit.
A novel security exploit allows malicious instructions, hidden as white text in Microsoft Word documents, to silently manipulate Microsoft 365 Copilot's output. The attack, disclosed by researcher Håkon Måløy, leverages cross-domain prompt injection by formatting hidden text that Copilot's underlying model reads.
This technique enables the AI to alter documents, such as halving financial figures, and then copy the injected prompt into the new document, disguised as source tracking. This allows the malicious document to act as a carrier, spreading through internal workflows without traditional malware.
Microsoft has implemented two mitigations since March 6, 2026, including a model upgrade. However, Måløy was able to reproduce the attack even after these fixes, indicating a persistent vulnerability class.
Microsoft acknowledges the findings and emphasizes a defense-in-depth strategy, continuously strengthening safeguards. They recommend users treat external content with caution, review AI-generated content, and install updates. Måløy suggests adding provenance metadata to generated documents for traceability.
