Bugbounty agent benchmark
I built a black box benchmark for bug bounty agents in a week and ran sixteen hundred cells across six models, seven harnesses and one browser. The largest effect I measured was not the harness or the browser: it was waiting a few weeks for the next model.