Futurion Solutions
TrustPrompt
Redact Personal Data in the Browser, Before a Prompt Is Sent
Detection runs on your machine. Personal data is redacted before the prompt is sent.Built for the clinics and law firms where a pasted prompt can carry patient or client data under the GDPR.
Works on the LLM Sites You Already Use
TrustPrompt runs on nine LLM web apps. Alongside the logos shown, it also covers Perplexity, Grok on X, and OpenEvidence — redaction happens in each site's composer before the message is sent.
What TrustPrompt Does
Personal Data Is Removed Before It Can Leave the Browser
Two things are always true: your prompt text is analyzed on your own machine, and detected values are replaced with placeholders before the message is sent.
Detection
A Privacy That Comes From Architecture
No detection backend exists
The model runs in the browser and remote loads are disabled, so there is nothing to send text to for inspection. Hosted AI gateways and proxies, by contrast, must receive the text in order to inspect it.
Nothing crosses a border
There is nothing to host and nothing that leaves the device. After install, detection works fully offline — which removes an entire class of data-processor and cross-border-transfer questions for regulated and public-sector buyers.
The third party never sees the real values
Redaction happens before the LLM receives the prompt, so the values are replaced before they leave the browser. There is no copy of the original text to log, retain, or train on downstream.
Deterministic regex with checksums
Structured data is matched by 40 built-in rules with checksum validation — IBAN (per-country length plus MOD-97-10), payment cards (Luhn), Spanish DNI/NIE/NIF/NUSS, passports, and secrets such as API keys, bearer tokens, and PEM private keys. These are exact and fast.
Transformer NER for the rest
Model finds the unstructured data no regex can catch reliably — names, addresses, organizations, and demographic and clinical attributes. Model exposes 54 entity types across Twelve languages.
One value, counted once
When both detectors fire on the same value — a regex DNI and a model identifier over the same digits — a cross-source dedupe step collapses them, so a value is never double-counted and the preview stays readable.
Local at Interactive Speed
A 560M-Parameter Model Running in the Browser, on Your GPU
Model Runs on Your Hardware
The runtime tries WebGPU with fp16, then WebGPU with q8, then WASM with q8. Backend selection is validated up front, so an incompatible GPU produces an explicit error rather than silently returning zero detections.
Why local is even possible here
Running a 560M-parameter transformer locally, at interactive latency, inside a browser extension is what WebGPU makes practical. Tools that offload detection to a server often do so because they cannot run a model of this class on the client.
Works without a qualifying GPU
The fast path needs a GPU exposing shader-f16 with around 512 MB of max buffer size. On machines without one, the WASM path keeps detection fully local at lower speed — it does not fall back to a server.
What Happens When You Press Send
Each prompt moves through the same sequence in the page, on your machine, before any text reaches the LLM.
Step 1: Block the Native Send
The site's own send is intercepted in the capture phase and held, so nothing leaves the page until detection and redaction have run.
Step 2: Built-In Rules and Your Watchlist
Deterministic, checksum-validated regex patterns run first, gated by the categories you or your administrator enabled, followed by your own custom watchlist terms.
Step 3: On-Device Model Detection
Each chunk of text is classified by the on-device NER model; overlapping detections are merged, aliases collapsed, and each entity is kept only if its score clears the configured threshold.
Step 4: False-Positive Suppression
A tiny on-device classifier, trained only from your own previous preview decisions, suppresses learned false positives. It stays on your machine.
Step 5: Preview
A panel lists what was found. You can dismiss an individual false positive or cancel the send entirely before anything leaves the page.
Step 6: Redact and Re-Send
Detected spans are replaced with typed placeholders, organization regex rules run last in an isolated worker with a timeout, and the redacted text is written back and sent.
Pricing
From a Free Single User to a Centrally Managed Estate
The free tier gives one person local redaction on all nine sites at no cost. Paid plans add shared configuration and your own detection rules for a team; the managed tier adds central policy. SMB prices are per organization, ex-VAT, billed annually.
Individual
One user, fully local
- Local redaction on all nine LLM sites
- User-owned on/off toggle
- Personal watchlist of custom terms
- On-device false-positive learning
- All seven categories on, fail-closed
Small Team
Up to 25 users
- Everything in Individual
- Shared sensitivity settings
- Custom regex for your own identifiers
- Employee, ticket, and patient ID rules
- Simpler rollout than full MDM
Business
Up to 100 users
- Everything in Small Team
- Shared configuration across the team
- Custom organization regex rules
- Priority support
Growth
Up to 250 users
- Everything in Business
- Larger team rollout
- Onboarding assistance
- A path to centrally managed policy
FAQs
TrustPrompt — Frequently Asked Questions
Straight answers for the clinics, firms, and teams deciding where their prompt text goes.
Is my prompt text sent anywhere for detection?
Which sites does TrustPrompt work on?
What does it detect?
What are the hardware requirements?
Which languages are supported?
How is it managed across an organization?
Does anything get logged?
What happens if detection fails?
How much control does the user keep?
Keep Personal Data Out of Public LLMs
Tell us about the sites your team uses and the identifiers you need to protect. We will show you how TrustPrompt fits, with detection running on your own machines.