How do you keep sensitive data out of prompts?

Summary of public sourcesWritten by our research agent from 7 public sources. Not first-hand experience.

Sensitive data reaches prompts mostly by copy and paste. A LayerX report found about 18% of enterprise employees paste data into AI tools. In the same report, 71.6% of AI access came through non-corporate accounts that sit outside company controls. Practitioners add a newer path: a coding agent pulled customer records from a production database during triage.

The approaches overlap across sources. Pick an approved provider under a contract that limits training use and retention, and block the rest. OpenAI's business terms, for example, say business data is not used for training by default and offer zero data retention on the API. Run models inside the company cloud or locally. Mask or tokenize data before it leaves. OWASP lists sanitization, input filtering, least-privilege access and user training. It also warns that a system prompt alone is not a reliable control. Tool settings help too. GitHub's content exclusion keeps chosen files away from Copilot, though its docs say it is not supported in agent mode. Practitioners say data loss rules miss a lot.

Still open: how to stop agents from reaching production data while still letting them debug.

Sources

• Hacker News: Employees regularly paste company secrets into ChatGPT
• Hacker News: Ask HN: How companies are protecting Claude Code from reading IP and PII data
• Survey: 77% of Employees Leak Data via ChatGPT, Report Finds
• Standard: LLM02:2025 Sensitive Information Disclosure
• Vendor docs: Content exclusion for GitHub Copilot
• Vendor docs: Business data privacy, security, and compliance | OpenAI
• Reddit: Ho do you let staff use AI without company data leaking out?