Detect
Find direct identifiers before text is sent. The prototype covers a limited set; measured results show what it misses.
// Icelandic data minimisation for AI
dmz masks kennitölur, names and account numbers before a prompt reaches a model, then restores real values in the answer. Real values stay on your device or in Iceland.
Draft a reply for Arna Jónsdóttir, kt. 140182-8858. Send to demo@example.com.
Draft a reply for ⟦NAME-1: XXXXX XXXXX⟧, kt. ⟦KT-2: XXXXX-XXXX⟧. Send to ⟦EMAIL-3: XXXXX@XXXXX.XXX⟧.
Hello , here is your reply.
Illustration of planned behaviour; actual markers also include a random session ID. All values are synthetic and test kennitölur use century digit 8. The working Shield prototype uses a manual side panel.
Staff keep the AI tools they use. The design keeps direct identifiers out of the prompt sent to a model.
Find direct identifiers before text is sent. The prototype covers a limited set; measured results show what it misses.
Every identifier receives a visible, distinct marker. A kennitala becomes ⟦KT-1: XXXXX-XXXX⟧.
When markers return unchanged, the matching real value is restored on device. Streaming restoration remains planned.
PHASE 1 DETECTION
Names and kennitölur do not follow English patterns. This prototype is a first step; a general name lexicon and inflection data are still missing.
Illustration of planned inflection detection, not current measured coverage.
Direct identifiers only. Context may still identify someone.
The Go core now runs in the browser as WASM and in Go tests. Portal and Gateway are planned to use the same core.
A manual side panel with the Go detector in WASM. It masks and restores on device. Automatic AI-site protection is next.
Explore this path →A workspace intended to route redacted text to local models or verified API routes.
Explore this path →A compatible API gateway with secret detection and metadata-only audit evidence.
Explore this path →FOR DPOs AND PROCUREMENT
Security claims should have evidence. We publish the prototype limits and will expand the measured corpus.
PUBLIC BENCHMARK
The selected 42-case corpus has 0 annotated identifiers in allowed outbound text, but a separate 18-case challenge set exposes 8 identifiers the prototype would let through. It is not ready for sensitive data. See results →
Markers show targets, not current performance.
30-day paid pilot · minimum 10 seats