Every prompt is a data transfer. When someone pastes text into ChatGPT, Claude, Gemini or Grok, that text leaves the device and lands with a third-party AI provider. LLM data privacy starts with one question: what is in the prompt before it is sent?
What leaks through prompts
The most common leaks are ordinary work:
- **Credentials:** API keys, passwords and connection strings inside pasted code or logs.
- **Client records:** names, account numbers, policy numbers and Social Security numbers in claim files or spreadsheets.
- **Private code and internal hostnames:** stack traces and configuration files shared for debugging.
- **Health information:** patient details copied from clinical notes.
Nobody intends harm. They have a deadline, and the AI tool is one paste away.
Why policies alone fail
Banning AI tools pushes the work to personal devices. Training helps, but it relies on everyone remembering, every time. Network proxies see traffic after it leaves the browser and add rollout friction. The most reliable control sits where the paste happens.
What prompt redaction is
Prompt redaction finds sensitive values in text and replaces them with placeholders before the text is submitted. The AI still gets enough context to help, but the secret itself never leaves the device.
A good redaction workflow:
1. **Runs locally.** Detection and masking happen on the device, so the raw prompt is never sent to another service for inspection. 2. **Works where people type.** It covers the AI tools and search engines people actually use in the browser. 3. **Shows the user what changed.** Coaching people in the moment builds better habits than silent blocking. 4. **Deploys centrally.** Security teams and MSPs need managed rollout rather than asking every user to install something.
A simple checklist for teams
- List the AI tools in use, including search engines with AI answers.
- Identify the data types that must never leave: credentials, client identifiers, health and financial records.
- Put a local masking control in the browser, not only a written policy.
- Roll it out through your existing management tools (Chrome Enterprise, Intune, Jamf, Group Policy or RMM).
- Keep reporting to counts and categories, so monitoring does not create a new store of sensitive data.
How Cloakworks handles this
Cloakworks masks passwords, private code and client records on the device before a prompt is sent to ChatGPT, Claude, Gemini or Grok. The AI provider only receives masked text, and the fleet dashboard stores counts and categories, not prompt content. See the [Security page](/security) for the data flow, the [deployment guide](/docs/deploy) for managed rollout, and [Cloakworks for MSPs](/cloakworks) if you protect clients in regulated industries. Questions are answered in the [FAQ](/faq).