ZONE A · USER DEVICE
Shield
Side-panel text↓Detect + redact (WASM)⇅Session map · device memory
// Trust and data boundaries
Where your data goes. Where it does not.
The Shield prototype processes locally in the browser. Portal and Gateway are designed for an Icelandic core that would mask before text goes to a model.
Data flow
ZONE B · ICELAND · PLANNED
Portal & Gateway
User · application↓Detect + redact + route⇅Encrypted session mapMetadata-only audit
ZONE C · MODEL PROVIDER
Receives redacted text
ChatGPT / Claude / Gemini webLocal model · Bedrock EUApproved API route
Context can still identify someone. Retention depends on the chosen route.
Boundaries by surface
SurfaceReal valuesProvider receivesRetention
ShieldOnly in device memory in the prototype.Redacted text under the AI site’s terms.Map is cleared on reset or when the side panel closes.
PortalPlanned Icelandic core and encrypted mapping.Planned: redacted text to a selected model.Retention and routes will be verified before launch.
GatewayPlanned Icelandic core and encrypted mapping.Planned: redacted text to an approved API.Planned metadata log without prompt bodies.
Evidence status
Data processing agreementDraft · counsel review pending
Sub-processorsPlanned list
Provider zero data retentionNot claimed
Bedrock EU retentionNot verified in account
ISO 27001Target · end 2028
Benchmark v042 selected + 18 challenge cases · 8 known misses
Limitations
×Cannot remove contextual clues such as jobs, dates or rare circumstances.
×Cannot promise 100% detection. Uncertain cases require review or blocking.
×Does not automatically make remaining text non-personal data.
×Does not detect documents, images or all Icelandic names in the prototype.
×Does not automatically intercept ChatGPT or Claude submissions yet.
Report a vulnerability: /.well-known/security.txt