Security & architecture
Trust the architecture, not just the promise
To protect your data, sitr has to handle it - so here is exactly what that means, and where the lines are drawn. Every point below is a property of how the product is built, not a slogan.
The AI provider receives masked text
Detection and masking happen before a prompt ever leaves for a model. Detection is statistical, so a value can be missed, and a value an employee chooses to send in the clear is sent as written. Providers receive tokenised text like [NAME_1]; the real values are restored only in your own employee's browser.
Self-hosted detection engine
The engine that finds sensitive data runs on private infrastructure with no public access - no third-party API sees unmasked text. Enterprise can self-host the entire stack, so raw text stays on your own network.
Encrypted token vault (AES-256-GCM)
Real values are encrypted in the app layer; the database only ever holds ciphertext. Keys are versioned and rotatable, and each value is cryptographically bound to its conversation - ciphertext copied anywhere else simply won't decrypt.
No cross-chat correlation
Tokens are scoped per conversation - the same person is a different token in every chat - so stored data can't be stitched back together into a profile of anyone.
Fails closed, by design
If the privacy layer is ever unreachable, the request simply stops - nothing is sent to a model and nothing is stored. Detection is not perfect: a value can be missed, which is why employees can review and add terms, and why we publish known limits.
Erasure on request
An operator-run erasure deletes a company's encrypted vault rows, after which the real values cannot be recovered from the masked text. There is no automatic retention schedule yet, and backups are covered by your agreement.
Your keys, your models, isolated
Each company brings its own provider credentials, encrypted and scoped to that tenant. Model traffic uses your own keys, never a shared one. Optional advisory features use a separate platform key and send masked text only. sitr never trains any model on your content.
Every send is reviewed and logged
Your employee can review what was detected before a prompt goes out. Detected values are masked by default; a value they explicitly choose to send in the clear is sent as written, and that choice is recorded in the audit log.
Written for your security team
Every claim on this page maps to the code that implements it. We'll share a full security-architecture document with your reviewers, so their due diligence starts from specifics - not marketing.
Built for peace of mind
The speed of AI, with far less exposure
Everything on this page describes how the product is built: selected sensitive values are masked before they reach a model, real values are encrypted at rest, and sitr never trains a model on your content. It reduces exposure; it does not remove it.
The technical foundation for the compliance work ahead of you:
Masked
text sent to AI models by default
Logged
as masked transcripts for audit
16+
AI providers, one interface
Any
industry or size
Give your team AI - and give yourself the audit trail
See sitr masking real prompts against your own data in a 20-minute demo.
getsitr.com · No commitment