Start with credentials
Passwords, private keys, access tokens, connection strings, recovery codes, and live API keys should never be included in a public AI prompt. Treat accidental disclosure like any other credential incident: revoke or rotate the secret, review access, and document what happened.
- Passwords and one-time codes
- API keys and bearer tokens
- Private SSH or TLS keys
- Database URLs containing usernames or passwords
- Session cookies and recovery codes
Limit personal and financial data
Avoid Social Security numbers, full card numbers, bank or routing details, tax identifiers, private addresses, and identifiable client financial records. Redaction should remove the sensitive value while keeping enough surrounding context to complete the task.
Protect company-confidential context
Confidential projects, acquisition plans, unreleased pricing, customer lists, contracts, and proprietary source code may not match a universal pattern. Companies need exact protected terms and workflow-specific rules rather than broad claims of perfect semantic detection.
When sensitive context is required
Use an approved enterprise environment with a reviewed data agreement, minimize the data, redact first, and confirm retention and training settings. If the company has not approved that workflow, pause and ask rather than guessing.