scrubbd

Local-first security log redactor. Pseudonymises logs in the browser - no accounts, no cookies by default, no analytics, no server-side processing.

Keep code identifiers neutral (redact, analyze, mapping, finding) - brand only in UI copy, package name, and docs. UI is English only for now (a French catalogue is prepared for a later language switch). Code, comments, tests and docs are in English.

Status: v0.1, work in progress. The analysis pipeline and the page work; see "Known limits" below for what is not done yet.

What it does

Paste or drop a log (plain text, JSON/NDJSON, CSV, CEF/LEEF, syslog, Windows XML, ...). scrubbd detects personal data and secrets (IPs, emails, accounts, hostnames, person names, phones, cards, tokens and passwords, ...), lets you keep or replace each value, then produces a cleaned file and a report. Replacements are consistent: the same value gets the same alias everywhere in the file. Secrets are replaced by [SECRET], never by a reversible alias. A final pass looks for original values left in the output.

Stack

• TypeScript (strict) + Vite
• UI without framework
• Vitest + Playwright
• Apache-2.0
• Cloudflare Pages headers (public/_headers) for CSP

Develop

npm install
npm run dev
npm run check          # type-check + unit/regression tests
npm run test:corpus    # the three log corpora (tests/corpus*)
npm run test:e2e       # Playwright: no network request after page load

Layout

src/core        detection, mapping, apply and verify pipeline (no DOM)
src/rules       field-name classification rules
src/ui          page and event handling
src/i18n        UI copy (English default + French)
src/worker      reserved for the Web Worker (not implemented yet)
tests           unit, regression, e2e and log corpora (see tests/README.md)
docs            decisions, specification, known limits

Principles

1. Zero network after page load
2. Structure before regex
3. Few precise detectors; low-precision ones opt-in
4. Transparent replacements; honest about limits

Known limits

scrubbd pseudonymises obvious values; it does not guarantee anonymity or GDPR compliance, and a manual review before sharing is always needed. Main open points:

• Analysis runs on the main thread (roughly 0.5-2 s per MB depending on the content), so large inputs freeze the page until the Web Worker lands.
• Bare tokens in free text are only caught for a few formats (GitHub, AWS access key id, JWT, ...); Slack, GitLab, Google, npm, OpenAI keys and AWS secret keys are missed unless they follow a key= / password: style label. No national ID number detector yet.
• Re-running on an already pseudonymised log can alias the aliases.
• Very deeply nested JSON (thousands of levels) can overflow the stack.
• Names with no contextual cue (free text, URL paths, file names) are only partly detected.

The per-case list of open decisions lives in docs/known-limits.md. Fixed product choices are in docs/decisions.md; the original specification is docs/specification.md.

Contributing: see CONTRIBUTING.md; security reports: see SECURITY.md.