I turned my prompt-injection benchmark into something you can rotate, scrub and click.

Ten jobs × ten attacks × five delivery channels = a 3D lattice of 500 real runs. Time is the fourth axis: hit play and every run lands in the order it actually ran. Ordinary agent on the left, BULWARK (the one I built) on the right.

Ordinary agent: 150 of 500 attacks landed. BULWARK: 0, with 499 of 500 jobs still done. The one refusal is a poisoned invoice, and the page tells you why.

Every number is recomputed from the rows in your browser. Add ?selftest=1 and it checks itself against the published bench.

Live: https://greenaidigital.com/tesseract/

Write-up: https://contra.com/p/UsEOkMoW-tesseract-500-attacks-on-my-ai-agent-replayed-in-4-d

If you can name an attack that should be in the lattice and is not, tell me and I will run it.