sunblaze-ucb/exploitgym ? reverse-engineered prompt

Build me a benchmark app for testing whether AI agents can turn real security bugs into working exploits.

I want a project that can set up a bunch of vulnerable tasks from different real world targets, run them in isolated Docker containers, and let an agent connect through a controller and firewall so it can try to solve each task safely. It should include scripts to download or prepare the task data, check that the setup is working, pull the needed container images, and then start the evaluation flow end to end.

Please make it easy to run an agent against a sample task list, and include clear setup and usage docs so someone can get it working locally. If you need current details for container images or tooling, look up the latest docs online. Also include a way to score submissions and keep the benchmark data organized so it can be updated over time.