Read-only first
The agent begins with bounded investigation tools. Access to a model is not authority to change a post.
Security
Observe broadly. Authorise narrowly.
Monitoring systems see sensitive infrastructure. An agent can amplify that visibility. Watchpost’s security model therefore separates what can be observed, inferred, recommended, approved, and executed.
Typed actions
No arbitrary model-authored shell execution. Each action has a schema, capability, post scope, parameter bounds, and audit contract.
Every state-changing operation — logins, posts, rules, pairing, approvals, actions and incidents — is recorded with the acting user's identity and is visible to administrators in the Audit view.
Untrusted telemetry
Logs, labels, service output, remote metadata, and integration payloads may contain prompt injection and are never treated as instructions.
Server-side enforcement
Every HTTP and WebSocket operation is authenticated and authorised on the server. Browser visibility is not a security boundary.
Visible transitions
Observation, recommendation, approval, execution, verification, and rollback remain distinguishable in storage and UI.
No safety claims
Watchpost is not a safety system or PLC interlock. Industrial control and safety authority remain outside the product boundary.
Design invariant
The agent can propose an action. Policy decides whether that action exists for this actor and post.
Security details will evolve with implementation. The project will publish guarantees only when they are backed by code, adversarial tests, and reproducible evidence.
Detailed security model