Security overview

This page is the public, operator-facing summary. Internal engineering detail lives in the repository security model and is not republished here.

Principles you can rely on

1. Tenant isolation is physical where it matters. Per-tenant databases; fail closed when identity or tenancy cannot be verified.
2. Secrets do not ambiently enter agent sandboxes. Credentials are injected at the egress boundary for the agent to use, not to read. Workspace drives hold a small, inventoried set of short-lived drive secrets outside the user-visible root.
3. Identity is verified, never asserted from the request body. Attribution and authorization are separate: shared sessions attribute messages; capability checks gate actions.
4. The model is an untrusted relay. Values that select whose data is touched must be something the model can only relay (for example signed refs), never mint.
5. Access is granted, never assumed. Services and workspaces start private; opening them is explicit.

What humans verify

Human review is part of the security and quality posture for judgment work. Approvals and corrections are attributed. External stakeholders only see what their grant opens.

Reporting issues

Contact support@psci-labs.com for security concerns with your tenant. Enterprise questionnaires and deeper posture reviews are handled through your design-partner or account channel.