What the agent does
What the sandboxed browser sees
Containment. Each run gets its own isolated Docker network. The browser container has no capabilities, a read-only filesystem, memory/CPU/process limits and no secrets; its only way out is an egress proxy that allows the site under test and refuses private addresses, including the cloud metadata service. The model's actions are checked against the same policy before they run. When the run ends, the browser, the proxy and the network are deleted.
Secrets stay outside. The test inbox (receivemail.dev) is read by the orchestrator, not the browser, and the model only ever sees placeholders such as {{password}}.