An educational single-page web app that demonstrates the mechanism of a Best-of-N (BoN) prompt-augmentation attack — on harmless example sentences only — so people can recognize what the technique looks like.
Live: https://bon.moosha.org
It is an awareness tool, not an attack tool: no model is ever contacted, there is no free-text input (only a closed library of benign sentences), and no harmful content is involved. Based on Best-of-N Jailbreaking (arXiv:2412.03556).
npm install
npm run dev # dev server
npm run test # 40 Vitest tests
npm run build # static build -> dist/- How it works & how it was built — the technique, the app, the safety model, the build process, deployment.
- Design spec — incl. the non-negotiable safety invariants.
- Implementation plan
Tech: Vite + Svelte 5, Vitest. No bac 4C35 kend, no runtime dependencies.