AgentProof is a production-style web app for authorized browser QA. Operators verify ownership of a target origin, generate a typed plan via Vultr Serverless Inference (`deepseek-v4.1-flash-1017`), approve only the steps they accept, and dispatch a run to a private sandbox broker. The broker is the sole Docker client: runners are non-root, read-only, capability-dropped, digest-pinned, and attached only to an allowlisted fixture network whose iptables policy permits TCP 80/443 to the owned app IP. After each task the container is removed. API keys never enter the sandbox. Challenge 1 Pattern B (browser) with Blast Radius Zero containment.
Category tags: