Theory · 8 min
Data leakage through AI assistants
Spot the ways company data leaves through AI tools, and the controls that see it.
The most common AI data loss is not an attack: staff paste source code, contracts or customer records into a public AI service to get help. Whether that service keeps or trains on the data depends on its terms and the account type.
Shadow AI is the use of AI services the company has not approved. Proxy and DNS logs show it; cloud access security brokers keep catalogues of generative AI apps and their risk.
Approved assistants leak too, through over-sharing: an assistant that can search all of SharePoint will find the file nobody meant to share. The fix is fixing permissions and labels, not blaming the assistant.
Microsoft Purview, for example, records Copilot interactions in its audit log, which lets an investigation ask what an assistant was asked and which files it touched.
Your turn: sort the evidence. Which practices increase the risk of AI data leakage, and which reduce it?
Do the interactive check Free, in the browser.
Sources
- OWASP Top 10 for LLM Applications 2025 OWASP GenAI Security Project
- Audit logs for Copilot and AI applications Microsoft Learn
- ATLAS data (technique IDs and names) MITRE