AI Privacy Firewall workflow
A deterministic, browser-side redaction layer that masks sensitive data before it ever leaves your machine — then safely rehydrates the AI's answer locally. Compliance teams get an audit log; engineers get a drop-in proxy that refuses to forward raw SSNs or card numbers.
Open this workflow or start your free 7-day trial.
What this workflow does
The Privacy Firewall sits between your team and any large language model. Sensitive entities — names, emails, phone numbers, SSNs, credit cards and your own custom terms — are detected and swapped for unique, reversible tokens (e.g. <EMAIL_1>). Only the sanitized prompt is sent upstream. When the model responds, the original values are re-inserted locally so the answer is still useful, without the raw data ever touching the network.
How to run it
1. Open the Privacy Firewall console.
2. Paste your system instructions and the sensitive data into the split inputs.
3. Add any project-specific custom redaction rules (internal codenames, account IDs).
4. Review the Safety Score dashboard and the action log of every masked item.
5. Send the sanitized prompt. Toggle Context Rehydration to control whether the AI's reply is de-tokenized locally.
Why it's safe by design
Redaction is deterministic and runs entirely client-side, so there is no "AI guessed wrong" failure mode. A second, server-side proxy acts as a last line of defense: it hard-blocks any request that still contains a raw SSN or credit-card pattern, so even a misconfigured client can't leak regulated data to the model.
Who it's for
Finance, Legal and Compliance teams that want to use AI on contracts, support tickets and customer records without violating data-handling policies, GDPR or PCI scope. Every masked field is logged, giving you an audit-grade record of exactly what was protected.
Frequently asked questions
Does my sensitive data ever reach the AI provider?
No. PII is tokenized in your browser before the request is sent, and a server-side proxy hard-blocks any prompt that still contains raw SSNs or card numbers.
Can I still read the AI's response normally?
Yes. With Context Rehydration enabled, the original values are re-inserted into the model's reply locally, so the answer reads naturally while the data stays protected in transit.
Can I add my own redaction rules?
Yes. You can add custom terms — internal codenames, account IDs, project names — that will be masked alongside the built-in PII detectors.