Prompt anonymizer
Pasting a prompt with real names, emails, and account numbers into an AI tool hands that data to a third party. The prompt anonymizer swaps each real entity for a realistic synthetic equivalent — and crucially, it does so consistently: the same person always becomes the same fake name. The model keeps all the context it needs to help you, but never sees the real values. A mapping lets you translate the answer back afterwards.
The problem with sending real data to AI tools
When you paste a prompt containing personal information into a commercial AI assistant, that data is transmitted over the internet to the provider’s servers. Depending on the provider’s terms and your settings, it may be logged, used for model improvement, reviewed by human annotators, or retained for a defined period. Even providers with strong privacy commitments hold the data in a system you do not control.
This is a meaningful risk when the prompt contains:
- Customer or employee names and contact details — personally identifiable information (PII) under GDPR, CCPA, and similar frameworks.
- Account numbers, reference codes, or partial payment data — which may carry regulatory sensitivity.
- Company-confidential details — contract terms, internal financials, HR information — where disclosure to a third party may breach agreements.
- Health or legal context — categories that carry enhanced protection under most privacy laws.
The straightforward mitigation is to replace the real values with plausible fictional equivalents before pasting. The AI model reasons about the structure and relationships in the prompt, not the literal identity of the entities, so the substitution preserves usefulness while removing the exposure.
How it works
The tool scans your text for entities — emails, phone numbers, two-word capitalized names, URLs, IP addresses, card-like numbers, and company names — and assigns each a stable synthetic replacement drawn from a fixed pool. A real value seen again reuses its earlier replacement, so relationships and references survive the substitution. You get the anonymized prompt plus a real-to-fake mapping.
For example, if your prompt contains “Jane Smith at [email protected] reviewed account #A-19283”, after anonymization it might read “Laura Barrett at [email protected] reviewed account #A-74912”. The AI can still help you draft a reply, summarize the situation, or analyse the pattern — without seeing Jane’s name or email.
Everything runs locally in the browser; nothing is uploaded.
Limitations and when to do more
The tool uses heuristic detection. It will catch the most common entity formats — email addresses with @ and a domain, phone number patterns, FirstName LastName sequences, credit-card-like digit strings — but it can miss unusual formats or over-match ordinary phrases. Always review the anonymized output before pasting.
For regulated data — health records, financial account data subject to PCI DSS, or any data covered by a formal de-identification requirement — a heuristic browser tool is a useful first step but is not a certified de-identification process. Follow your organisation’s approved redaction procedures for data that carries legal obligations. This tool is most useful for the large category of sensitive-but-not-regulated prompts where good privacy hygiene is the right call.