The security model, in full
Ridge exists because a shared API key handed to an entire team has no identity, no least privilege, and no audit trail behind it. This page is the detailed version of the posture on the homepage — what Ridge actually enforces, and what it deliberately doesn't claim.
Authorization: service × tier × client scope
Every tool belongs to a service and a tier. Tools that touch a specific client — an Autotask company, a Hudu company, a Conduit agent's site — are additionally scoped, and a grant lists which client codes it covers, or every one. Grants attach to a team or an individual member; a person's effective permissions are the union of both.
| Tier | What it allows |
|---|---|
| read | Look things up — search tickets, list assets, read an audit log entry. |
| write | Change non-destructive state — create a ticket, add a note, log time. |
| destructive | Held for human approval before it runs — reboot a device, uninstall an agent, run a command on an endpoint. |
| reveal | A separate, non-nesting axis. A destructive grant confers no reveal, and a reveal grant confers no read or write — disclosing a secret is a different kind of risk than changing state, so it's always granted explicitly and rate-limited. |
The tool list each person sees is filtered to what they're allowed to call — nobody sees a tool they can't use. A revoked grant takes effect immediately, not at the next sign-in.
What the gateway enforces
- OAuth 2.1 with PKCE for every client sign-in.
- Credentials are encrypted at rest with AES-256-GCM envelope encryption, using per-organization keys, with optional bring-your-own-key.
- Every request is validated before authorization runs.
- Audit records redact known secret fields before they're written.
- Upstream text returned to a model is treated as plain data, never as instructions.
- Ridge itself exposes no code-execution tool. It only ever calls the tools of the integrations you've connected, each with its own tier.
- Console admin actions — grants, connection changes, approval decisions — are audited too, the same as a tool call.
Destructive actions wait for a human
A destructive call doesn't execute immediately — it creates a pending approval. Someone with approval rights decides it from the console or right in the same chat conversation, and the request expires if nobody does. The decision itself is audited, the same as the call it gates.
What we don't claim
We don't hold SOC 2, ISO 27001, HIPAA, or any other compliance certification today, and we don't publish customer names or usage metrics. Ask us where we are on the compliance roadmap.