Skip to content

ADR-064: Range guests get default-on, keyless model access

Status

Accepted for #2275, 2026-09-19. Supersedes ADR-059's "no-service-account guest default" decision (ADR-059-R3) for GCP. Restores the default-on model-access enablement removed by the ADR-059 direct guest-credential retirement, without reintroducing guest-held provider keys.

Context

ADR-059 introduced the deployment-owned model broker and, in migration, retired the direct guest model-access path: the range guest Vertex service-account role and the AWS guest Bedrock authority were removed, and both GCE plan builders set attach_service_account: False.

The broker is the right control where mandatory spend, rate and request enforcement is required, but it is not turnkey. Enabling it requires operator-supplied TLS (broker and control secrets), a CA trust bundle, a private VIP and DNS hostname, a KMS fingerprint signing key, a v3 priced catalog, and a provider inventory, and it is disabled by default. The net effect of the retirement was that a fresh tenant has no model access at all until a heavyweight broker bring-up is completed, and range model-access enablement became coupled to whether a specific scenario declared a model need.

That coupling is wrong. Whether a scenario uses a model is independent of whether a deployment's ranges are able to reach one. Most deployments expect their own project's Vertex (or AWS Bedrock) to be reachable by ranges by default, with only per-model enablement in the provider console left as a manual step, the posture that existed before the retirement.

Decision

Range guests receive a default-on, keyless model-invocation identity, decoupled from scenario model-need.

On GCP, the range host service account carries a predict-only Vertex custom role (aiplatform.endpoints.predict) in the platform project (same-project by default), and the provisioner attaches it to range guests through Workload Identity. The guest obtains short-lived tokens from the metadata server; no key material is created, minted, or delivered to the guest. This resolves the original ADR-059 concern (Vertex keys placed in the guest, many keys authenticating one principal) while restoring reachability. aiplatform.googleapis.com is enabled by default. Enabling a specific model (for example Claude in Vertex Model Garden) remains the only manual, provider-console step.

On AWS, the range instance role mirrors this with default-on Bedrock invoke (bedrock:InvokeModel, bedrock:InvokeModelWithResponseStream); enabling a specific model in the Bedrock console remains manual. AWS is delivered separately and requires its own qualification, per ADR-059.

Direct default-on access and the ADR-059 broker are mutually exclusive per range. When the broker is the guest model path (MODEL_BROKER_GUEST_VIP set), the provisioner leaves guests identity-less and model access crosses the broker only, preserving its source-preserving keyless isolated egress. When no broker guest path is configured, the keyless guest identity is attached.

Alternatives

Alternative Disposition
Keep model access broker-only, off by default Reject: leaves every fresh tenant with no model access and couples enablement to a heavyweight per-tenant TLS/DNS/KMS/catalog bring-up.
Restore per-range Vertex keys (the retired path) Reject: reintroduces guest-held key material and the 10-key-per-principal / 24-slot limit. Keyless Workload Identity attachment gives the same reachability without keys.
Make the broker turnkey and default-on Not selected now: the broker's required TLS, CA, VIP/DNS, fingerprint KMS and priced catalog cannot be sensibly auto-defaulted. The strategic turnkey path is the AI-gateway work (milestone: AI Gateway, #2274).

Consequences

Range guests reach approved models by default with no key material and least privilege (predict-only invocation plus telemetry writes). A compromised guest can call approved models directly under this identity: direct-path invocations are not subject to the broker's mandatory request/spend/rate enforcement. Deployments that require that enforcement enable the broker, which switches ranges to the mediated, identity-less path. This is the deliberate tradeoff: reachable by default for the common case, broker-mediated where enforcement is required.

Network reachability (Private Google Access or external egress) is the operator's unchanged range egress posture and is orthogonal to identity: the old key path needed the same reachability.

The workload IAM scope guard is unchanged. The range host identity still holds no project storage or secret roles; only predict-only model invocation and logging/monitoring writes are added, so participant-reachable guests gain no project storage, secret, or cross-tenant access.

Audit telemetry clarification (#2535)

The GCP range telemetry preflight records the intended default-on network and Vertex Data Access logging boundary. Provider audit configuration belongs to deployment Terraform in the model target project; guest subnet/NAT settings belong to their existing infrastructure owner. Logging grants no new guest model permission, project IAM authority or log-read access, and does not enforce spending or identify a human behind a shared guest principal. Cloud Logging routing/retention remains project-operator policy. The broker stays parked under ADR-068. This clarification is design guidance; rendered configuration and live record evidence remain #2535 delivery work.