Skip to content

PromptWard 0.11.1

Latest
Latest

Choose a tag to compare

View all tags
@gduplessy gduplessy released this 18 Jul 16:22
v0.11.1
09c7531

PromptWard 0.11.1

Local-first prompt PII guard for ChatGPT, Claude, Gemini, Perplexity, and Mistral. Detection and redaction run entirely on-device via a local ONNX model — no prompt text, placeholder map, or telemetry ever leaves your machine.

This is the first release since 0.9.3 and includes a lot of fixes (see CHANGELOG.md for the full history). The two that matter most:

Install

  1. Download promptward-extension.zip below and unzip it somewhere permanent (don't delete the folder afterward — Chrome loads the extension from it).
  2. Go to chrome://extensions, enable Developer mode (top right).
  3. Click Load unpacked and select the unzipped folder.
  4. Visit a supported AI chat site and send a prompt containing PII. PromptWard shows a redaction review before it goes out, and auto-sends the redacted version after a 5-second countdown (or click Send original to opt out, or click anywhere on the modal to take manual control).

The local model loads automatically on first send — no need to open the side panel first.

What's new since 0.9.3

Critical fix — protection now actually works on ChatGPT, Perplexity, and Claude.
In 0.9.3 the content script resolved the wrong editor element on rich-text composers (Lexical/ProseMirror), read empty text, and let sends through without ever running detection. No modal appeared; prompts went out unredacted. Editor lookup now ranks all editable candidates and picks the one containing your text. Verified live on chatgpt.com and perplexity.ai.

The auto-send countdown no longer vanishes. The 5-second countdown was cancelling instantly on incidental mouse movement (the cursor drifting over the modal after your click). It now only cancels on a deliberate click or keypress.

Smaller and faster.

  • Packaged size dropped from ~93 MB to ~24 MB — only the ONNX Runtime WASM variant the runtime actually loads is shipped.
  • Debug logging moved off the send critical path — every send was paying serialized latency for diagnostics nobody sees.

Reliability.

  • Enter during IME composition (CJK input methods) no longer triggers a premature send.
  • Worker/protect requests can no longer hang indefinitely — they now time out and fail closed (nothing sent) instead of stalling forever.
  • Concurrent sends no longer race the model document (previously: a spurious "Prompt blocked" error on wake).
  • Custom-domain input is validated before permission is requested, so one bad entry can't silently break protection on all custom domains.
  • Retry results now route through the review modal like the initial attempt.
  • A per-page-load crash that silently disabled diagnostics is fixed.

Removed. The side panel's "WebGPU" toggle, which was labeled "Reserved opt-in" and did nothing, is gone.

See PRIVACY.md and NOTICE for data handling and third-party model attribution (Rampart model, ONNX Runtime Web).

Assets 3
Loading