Problem
When I ask an AI assistant for help, I often paste logs, config files or terminal output. That text can contain API keys, passwords or tokens. Secret scanners protect code repositories, but nothing protected the clipboard, the moment before text leaves the machine.
What I built
Redactor is a Windows desktop app and a command-line tool, scrub, for Windows, Linux and macOS. You paste text in, and it returns the same text with every secret replaced by a label like [REDACTED: Password #1]. The same value always gets the same number, so the assistant can still see that two lines use the same password. A global hotkey cleans the clipboard in place, and scrub can block a commit that contains secrets.
Everything runs on your machine. There is no network code, and the build fails if an HTTP library ever enters the dependencies.
How it works
One Rust engine serves both the app and the command line:
- It removes invisible characters that could hide a secret, and keeps a map back to the original text.
- Three detectors run: 181 rules for known token formats, a context engine that reads key names in JSON, YAML, XML, shell, SQL and plain text in five languages, and a small trained classifier for random-looking tokens.
- It looks inside base64, hex, percent-encoded and JSON-escaped text, and finds every other copy of a secret, encoded copies included.
- Each value gets one label, and the output is scanned again before it is called safe.
My part
I led the project. I set the requirements and the security rules, wrote the threat model, made the architecture decisions (Rust instead of the Python prototype, Tauri instead of Electron, no network code enforced by the build) and decided what shipped. The code was written by AI coding agents working in parallel under my direction. Every change had to pass the tests, an accuracy corpus and a review before it was merged.
Result
Version 2.0 shipped on 7 November 2026. Version 2.1 followed on 11 November, after six reviewers audited my own release and found leaks, which I then fixed. It comes as a 2.9 MB installer, a portable app and the command-line tool for four platforms.
On my labelled test corpus it finds all 811 planted secrets in 92 files, with no false alarms in 62 clean files. I tuned the detectors against that corpus, so it proves the known cases and says nothing about unknown ones. The threat model lists seven shapes that still get through, each with a workaround. 895 Rust tests and 183 UI tests run in CI on Windows and Linux. Releases are built only on GitHub’s servers, with checksums and build provenance attestations that anyone can verify.
What I’d do next
Sign the Windows files through the free SignPath Foundation program, bring the desktop app to macOS and Linux, and close the known gaps without adding false alarms.