Preview-first protection for trusted AI destinations. Individual and team deployments today; enterprise controls available through a tailored implementation.

Review Sensitive Data Before AI

A browser-extension prototype that checks pasted text locally using narrow rules and a compact MiniLM detector.
Review every candidate, choose Mask or Keep, then insert the reviewed text into a trusted AI destination.
Preview-first because the measured automatic-masking precision did not meet the 98% release gate.

Detect exact spans

MiniLM handles contextual spans
Rules handle narrow URLs and secrets

Review before masking

Every candidate appears in a preview
You decide what to Mask or Keep

Preserve the question

Remove incidental private detail
Keep public context needed for the answer

How it works

StepWhat happens
Paste interceptionThe extension listens for text pasted into an allowlisted editor.
Local analysisPackaged MiniLM INT8, tokenizer, ONNX Runtime Web WASM, and deterministic rules run locally.
Exact reconstructionBIOES decoding returns candidates to their original character offsets, including overlapping long-text windows.
User previewEach candidate stays behind Mask or Keep review; the extension does not silently rewrite the prompt.
Reviewed insertionThe selected result is inserted into the same editor. Cancellation makes no change.
Failure behaviorIf the scanner, model, or WASM fails, the original paste is restored so user work is not blocked. That failed paste is not protected.

Measured evidence

ResultEvidence boundary
0.886 exact-span F1MiniLM hybrid on a frozen 1,000-case synthetic/public detector test.
95.8% auto-mask precisionBelow the provisional 98% gate, so preview review remains mandatory.
About 196 ms warm p95Controlled 1K-character paste in retained Brave measurements on the development Mac.
31 runtime testsPlus retained browser evidence for normal, corrupt-model, missing-WASM, and unsupported-WebAssembly cases.

Read the frozen baseline, policy regression, runtime benchmark, and security audit.

Preserve useful context

Original prompt

My cousin Ananya studies at DPS Indiranagar. Is that school any good?

Reviewed prompt

My cousin [PERSON] studies at DPS Indiranagar. Is that school any good?

The private name is unnecessary. The school is the query anchor and remains available to the AI tool.

Questions & answers