Trust · Security architecture
One governed door, with clear boundaries.
IntSen sits between your applications and your approved model providers. Every governed call is authenticated, authorized and checked before it leaves—and recorded on the way back.
Trust boundaries
Three zones, one path between them
Your environment
Apps and agents
- Run on your infrastructure and frameworks
- Hold only an IntSen OAuth client
- Execute their own tools
IntSen
AI Gateway and governance services
- Authenticate and authorize every call
- Run guardrails and budget checks
- Hold provider keys; inject per call
- Record requests, verdicts, turns and scores
Model provider
Approved models
- Receives only allowed calls
- Sees no IntSen headers
- OpenAI in the first release
Authorization
Every call passes the same decision, in order
If any check fails, the call is refused with a specific code before it reaches the provider.
- 1A verified issuer and client id are present on the token
- 2The call names a registration (x-intsen-registration-id)
- 3The registration exists, is ready and is not quarantined, in the right environment
- 4The client is bound to that registration
- 5The provider is granted to the registration
- 6At least one model is usable: active in the catalog and cleared for the data class
- 7The endpoint and the requested model are allowed
Controls
Security controls at a glance
- Authentication
- OAuth 2.0 client credentials per application and agent. Team identity providers can be registered and validated.
- Authorization
- Policy decision point in the request path; fails closed; every denial returns a code and is logged.
- Model access
- Models come from a curated catalog with status and data-class clearance; registrations allow-list models per provider.
- Input validation
- Oversized and invalid requests are refused before they reach the provider.
- Policy enforcement
- Guardrails and budgets run in the request path; task-level checks after the turn.
- Credential handling
- Provider keys are held by the gateway and injected per call; never given to applications.
- Header hygiene
- All x-intsen-* headers are removed before the provider call.
- Change control
- Resources are declarative; use cases can require approval; changes keep history; deletes are soft and restorable.
- Abuse monitoring
- Usage monitors flag unusual activity on a use case or agent to its owners.
Review the architecture with us.
Bring your security architects; we'll walk through trust boundaries, data flows and failure behavior.