Microsoft Copilot for Word copies hidden prompts into new docs

Microsoft Copilot for Word was found to copy hidden prompts—metadata that records the AI’s internal instructions—into newly created documents, a behavior that could expose user interactions to unintended parties and spark regulatory attention [1].

Hidden prompt leakage in Word

The issue emerged as Copilot generated a new Word file and unintentionally embedded the prompt that guided its response. Because the prompt is not visible in the document body, users may assume the file is clean while the hidden data persists. Microsoft’s broader Copilot rollout already includes multi‑step, app‑native actions across Word, Excel and PowerPoint, but this new finding highlights a gap in data sanitisation [2].

Implications for users and the market

For enterprise customers, the hidden prompt could reveal internal queries or proprietary language, contravening data‑privacy policies that require explicit control over metadata. Competitors such as Google’s Gemini and Anthropic’s Claude have emphasized “clean output” as a differentiator, so any perceived lapse in Microsoft’s handling may affect adoption decisions. The incident also aligns with broader scrutiny of AI assistants that retain conversation history for up to 18 months, as noted in Microsoft’s privacy disclosures [1].

What to watch

• Microsoft response timeline – Look for an official patch or guidance from Microsoft addressing the hidden prompt issue.
• Regulatory scrutiny – Monitor statements from data‑privacy regulators that may reference Copilot’s metadata handling.
• Competitor positioning – Watch for marketing moves by Google or Anthropic that highlight stricter data‑output controls.

The discovery underscores the tension between powerful AI assistance and the need for transparent data practices, leaving open whether Microsoft will adjust its Copilot architecture to eliminate hidden prompt leakage.