sovereign-agent

Threat model

The current operational security policy is SECURITY.md. This page expands the model behind those boundaries for readers of the source.

Assets

Untrusted inputs

Provider output, prompts, deliverable paths, workspace contents, environment variables, stale processes, CLI arguments, proof-pack manifests, and direct filesystem changes are untrusted. A model’s confidence or a subprocess exit code is never authorization or proof of success.

Enforced controls

Residual risk

Explicit non-goals

The package is not a secrets manager, distributed scheduler, web control plane, general workflow engine, container orchestrator, or guarantee of autonomous operation without human governance. See non-goals.md.