Sandbox Tester running in a Docker container with fewer available command families is a proof of concept for a very practical question: if an AI agent is put in a container, what can it still do? The agent runs a broad set of capability checks and reports which actions are allowed, denied, unavailable, or broken. That makes the sandbox less of a comforting diagram and more of a receipt.

This version focuses on the execution-control profile. Earlier Docker hardening limited file access, network egress, and ambient services. This step asks a narrower question: what happens when needless command families are disabled? Package managers, source-control tools, admin commands, namespace tools, extra shells, and process helpers are blocked while Python, Playwright, Chromium, and the OpenAI API test still run.

The result is not a claim that the container is magically safe. Sensible people do not say that near computers. It does show that removing commands changes the agent’s practical reach.