Malicious instructions embedded in Word documents can compromise Microsoft Copilot and silently alter files without users knowing, according to security researcher Håkon Måløy. When these hidden directives are processed by the AI, they modify the output Copilot generates and propagate themselves to new documents created from the infected file.
Microsoft delayed public disclosure of the vulnerability twice. After nearly six months passed since Måløy's initial report, the researcher decided to go public with the findings. The key takeaway: treat any documents from external sources as potentially untrustworthy, especially when using them with AI tools like Copilot.
