Universal AI ↔ Enterprise Data control plane · 31 backends · HMAC-audited Read the whitepaper →
AegisAI
AI on GCP · Workload Identity Federation

AI agents on GCP without service account sprawl.

AegisAI uses Workload Identity Federation to propagate the calling user's identity to BigQuery, Vertex AI, Cloud SQL, and Spanner — without distributing service account keys, without long-lived credentials, and without breaking organization-policy guardrails.

Build status
GCP integration
as of June 2026

✓ LIVE: BigQuery with Workload Identity Federation. testIamPermissions check framework. Service account JSON-free auth path.

⧉ EARLY ACCESS · Q3 2026: Cloud SQL IAM database auth, BigQuery row access policy enforcement under per-user federated identity, VPC Service Controls perimeter preservation

⧉ ROADMAP · Q4 2026: Vertex AI Search per-user document scoping, Agent Builder tool-call identity propagation, Gemini grounding ACL honoring, Spanner FGAC per principal, AlloyDB column-level grants

WIF + STS token exchange is built once and reused per GCP service. Each new data plane integration is ~5 engineer-days against a customer GCP sandbox.

The GCP-specific problem

Google built the most modern federation story in cloud. Naive AI integration uses the oldest pattern in cloud.

Workload Identity Federation, VPC Service Controls, Organization Policies, BigQuery row access policies, Vertex AI grounding — Google's identity tooling is genuinely best-in-class. The naive AI assistant pattern: a service account JSON key in an environment variable. Every modern control above gets bypassed.

Service account key sprawl

Each AI integration creates a service account. Each service account gets a JSON key. Each key lives in some .env or secrets manager somewhere. Compromised key = full data access. Google's own docs say "don't use service account keys" but the AI tutorials all do.

VPC Service Controls perimeter

VPC SC perimeters protect BigQuery datasets from exfiltration. A service account inside the perimeter that calls out to an external AI vendor = perimeter break. Per-user identity preserves the perimeter semantics.

BigQuery row access policies

BigQuery row access policies (RAPs) evaluate SESSION_USER(). A service account makes that meaningless. Per-user identity propagation makes RAPs work as designed.

How AegisAI integrates with GCP

WIF all the way down. No service account keys.

AegisAI exchanges the user's OIDC token for a Google STS token via Workload Identity Federation. The Google token impersonates the user identity. BigQuery, Vertex AI, and Cloud SQL evaluate IAM against the actual user. No keys distributed, no long-lived credentials anywhere.

BigQuery · dataset ACLs · row access policies

Dataset-level IAM, table-level IAM, row access policies, column-level masking — all evaluate against the calling user. BigQuery Audit Logs in Cloud Audit Logs attribute queries to actual users for SOC 2 and FedRAMP audits.

  • Per-user dataset and table grants enforced natively
  • Row access policies fire correctly because SESSION_USER() is real
  • Column-level masking respects per-user policy tags
  • BigQuery slot allocation per user for cost attribution
  • BI Engine reservation usage tracked per analyst

Vertex AI · Search · Agent Builder · Gemini

Vertex AI Search uses per-user OAuth scopes. Agent Builder and Gemini grounding can call BigQuery / Cloud SQL under user identity instead of the Vertex AI service account. Per-user quotas, content filter logs, and prompt audit all attribute correctly.

  • Vertex AI Search per-user document permissions
  • Agent Builder tool calls propagate user identity to backend
  • Gemini grounding sources respect dataset ACLs per user
  • Model Garden access controls per user, not per shared SP
  • Vertex AI Pipelines step execution attributed to triggering user

Cloud SQL · Spanner · AlloyDB

IAM database authentication for Cloud SQL maps the user's federated identity to a Postgres or MySQL principal. Spanner fine-grained access control and AlloyDB column-level grants fire against the actual user.

  • Cloud SQL IAM authentication (no postgres superuser everywhere)
  • Spanner Fine-Grained Access Control per principal
  • AlloyDB column-level grants enforced
  • Query Insights groups by actual user identity

Organization Policies · VPC SC · Cloud Audit Logs

Organization-level policy enforcement still applies. VPC Service Controls perimeters protect dataset egress at the organizational layer. Cloud Audit Logs answer "who accessed this dataset?" with the actual user, not the integration service account.

  • Org Policy constraints applied per request
  • VPC SC perimeter semantics preserved (no service account egress hole)
  • Cloud Audit Logs attribute access to actual user federated identity
  • Per-folder / per-project policy boundaries respected
Architecture · GCP-specific flow

From the user's IdP to BigQuery, no keys distributed.

External IdP
Okta / Auth0 / ADFS
AI Agent
Gemini / Vertex / Claude
AegisAI Gateway
OIDC validation
↓ Workload Identity Federation ↓
Google STS
Token exchange API
WIF principal
Bound to user identity
VPC SC perimeter
Org boundary check
↓ Per-user data plane call ↓
BigQuery
RAP + dataset ACL
Vertex AI
Per-user quota / log
Cloud SQL
IAM database auth
Why this matters for GCP shops

Five things you stop worrying about.

Zero service account keys

Workload Identity Federation eliminates the JSON-key-in-environment-variable anti-pattern. Compromised AI tooling can no longer leak persistent credentials.

VPC SC perimeter holds

Exfiltration protections at the org perimeter remain enforceable. AI access doesn't become a perimeter bypass.

BigQuery RAPs work as designed

Row access policies based on SESSION_USER() finally do what your data engineering team intended.

Cloud Audit Logs are actionable

Forensic investigations get real user identities, not a sea of "service-account@..." entries.

Ready when you are

One identity. Every cloud. Every AI agent.

30-minute architecture call. We open the operator console and run real queries through your stack — AWS, Azure, GCP, or all three. You see the audit chain tick up in real time.