Security 6 - defense in depth beyond the app
Digitorn's security model has two layers, both of which run in-process and are described in full in Security Architecture:
tools.capabilities- the seven gates that decide which actions an agent can call (grant / approve / deny / risk caps). See tutorials 1-2.security.behavior- the behaviour engine, which intercepts patterns of tool use even for granted actions. See tutorial 3.
Both run as regular Go code in the same process as the agent loop. A bug in either, or an exploit chain that defeats both, reaches the underlying OS without further checks. For a process/kernel isolation boundary on top of that, run the daemon itself inside a container, VM, or your own sandboxing wrapper.
runtime.workdir_mode (see
App Configuration) resolves and
creates the directory an agent's filesystem tools default to.