npm package · live · 2026-10
part of foxmate
foxbench
a fixed task suite that scores browser agents in firefox
the problem
it is hard to compare browser agents. an agent's own report of success is not proof, and a page can hide an attack.
what i built
foxbench serves four mock sites on your machine and gives an agent 13 tasks. an oracle reads the server state to decide if each task passed. four tasks hide a prompt injection, and the score shows how many attacks the agent blocked.