When this page loaded, it ran all 209 prompts from the project's two red-team test sets through the same engine you are about to use. The first set has 50 secrets and personal-data cases, including traps that should not trip it, like Git hashes and invalid card numbers. The second has 159 prompt-injection cases in nine languages, including innocent sentences built to trip it, like "don't ignore the instructions on the label". The browser engine runs the exact pattern table the Python engine exports, and is checked against it, finding by finding, on every commit.
Gatekeep checks every prompt on its way to the Claude API. Pick one, send it, and see what happens. It all runs in your browser.
Measured on public datasets it was never tuned on: the patterns were written against the training splits; these are the held-out test splits plus two sets never used at all. Each bar shows the 95% confidence interval; the tick is the measured rate.
Every decision gets a row. The prompt text itself never does.
Same columns as the proxy's real audit table. Gatekeep keeps a SHA-256 fingerprint of each prompt, what kinds of things it found, where the request went, and how long the check took. The raw prompt is never written down, even here.