Deployment security and fail-closed
TokenVeil is self-hosted, so the strongest guarantees come from how you deploy it, not just from the software. This page covers the security posture and how to make a deployment fail-closed, which is a common requirement for regulated environments.
Where your data lives
Section titled “Where your data lives”The mapping between real values and tokens never leaves your instance. It is encrypted at rest and is addressable only through an opaque session id. API access is authenticated per key, with expiry, quota, rate limiting and revocation, and every transformation is audited by type and count, never by real value.
Making a deployment fail-closed
Section titled “Making a deployment fail-closed”A deployment is fail-closed when sensitive data cannot reach a provider even if something goes wrong. Two conditions achieve that:
- Lock network egress. Configure the network so that only TokenVeil can reach the model providers. The client has no direct route to the provider, so it cannot bypass the anonymization layer by calling the provider itself.
- Block on failure. Configure TokenVeil to reject a request if anonymization does not complete, rather than passing the original content through. A failure becomes a blocked request, not a leak.
With both in place, there is no path where original content reaches the provider without passing through anonymization.
Scope boundary to be explicit about
Section titled “Scope boundary to be explicit about”The anonymization layer covers the content sent to the model. Clients such as coding agents also make non-model network calls (telemetry, update checks) that do not carry your sensitive context. If your policy requires zero direct provider or third-party egress at all, those calls belong in the same network egress policy as point 1 above. Being explicit about this boundary is part of a fail-closed design, not an afterthought.
Transparent gateway (roadmap)
Section titled “Transparent gateway (roadmap)”A transparent mandatory gateway for coding agents such as Claude Code and Codex, where the tool is pointed at TokenVeil as its model endpoint and the full model-bound context is anonymized on the wire with no manual step, is in development. See the roadmap. Today the same protection is achieved through the API and a thin integration; the transparent drop-in is the next step and is a natural pilot deliverable.