How to use ChatGPT at work without leaking sensitive data
The method to use ChatGPT, Claude or Gemini at work without exposing sensitive data: anonymize the prompt before sending, then restore the answer locally.
To use ChatGPT, Claude or Gemini at work without leaking sensitive data, the principle is simple: replace every sensitive value with a reversible token BEFORE sending the prompt, then restore the original values in the answer. The assistant never sees the real name, email, IBAN or API key — it works on tokens like [PERSON_NAME_001], and you get back a complete, readable text in your browser.
Why pasting sensitive text into an AI assistant is risky
When you paste a customer email, an HR record or a snippet of code into a consumer assistant, that content leaves your machine and travels to a third-party provider. Depending on the contract and settings, it may be stored, processed outside the EU, or reused. For personal data or a company secret, that is a transfer most security policies forbid.
- Personal data: names, emails, phone numbers, addresses, national IDs.
- Financial data: IBANs, bank details, salary amounts, tax identifiers.
- Technical secrets: API keys, access tokens, login credentials.
- Strategic data: client names, meeting markers, internal URLs.
The method: anonymize, send, restore
Rather than banning AI — which pushes teams toward unmonitored usage — you govern the flow. Three steps, two of them invisible to the user:
- 1Detection: an engine spots sensitive data in your text (names, emails, IBANs, API keys…).
- 2Tokenization: each value is replaced with a neutral, reversible token kept in local memory.
- 3Restoration: the assistant's answer is de-tokenized in your browser — you read a complete text, and the AI never saw the original.
What to check before adopting a tool
- Is processing truly local (in the browser) or server-side?
- Does detector coverage go beyond emails and card numbers?
- Is restoration reliable, including on long answers?
- Can you add your own rules (project names, internal codes)?
ONYRI Sanitize answers all four: the tokenization engine runs 100% in the browser, covers six families of sensitive data, restores answers segment by segment, and accepts custom rules per workspace.
Set a simple team rule
The best policy fits in one sentence: “No sensitive data leaves in clear text to an AI assistant.” Give teams a tool that makes the rule easy to follow, rather than a ban that invites workarounds. Document the categories to protect, enable the matching detectors, and let everyone work with AI confidently.
Frequently asked questions
- Does ChatGPT keep what I paste?
- Depending on the plan and settings, content may be stored and processed by the provider. The reliable defense is to never send the data in clear text: you send already-anonymized text.
- Does anonymization degrade the AI's answer?
- No. The assistant reasons over consistent tokens (the same name maps to the same token). Text structure is preserved, and the answer is restored with the real values in your browser.
- Do I need a dedicated tool or can I do it by hand?
- By hand, you always miss a value and can't reliably restore. An automatic engine detects, replaces and restores reliably — which is what makes the practice sustainable day to day.
Sources & references
Keep your sensitive data in your browser
ONYRI Sanitize detects and masks your sensitive data before it reaches the AI, then restores the answer — from names to API keys.
Anonymize my prompt