Google just closed a decades-old convenience loophole for its most powerful model APIs. Starting mid‑June 2026 Google began rejecting "unrestricted" API keys for Gemini and will reject all "standard" API keys by September 2026 — replacing them with service-account-backed auth keys that are identity-bound and scoped by default. This is not a paperwork change; it rewrites the trust boundary for every CI job, serverless function, and edge service that talks to Gemini.
This is the right call. Unrestricted API keys were a single-point blast radius: leak a key and an attacker gets full model access. Service-account-backed auth keys force you to think in identities, scopes, and short lifetimes. But it also means a painful migration for teams that treated API keys as long-lived config values stuffed into env vars or baked into container images.
The immediate impact
- Enforcement timeline is concrete: unrestricted keys rejected starting mid‑June 2026; vendors and platform automation relying on standard keys must be migrated by September 2026.
- Google’s older "generate"/"generateText" style endpoints are being supplanted by consolidated Responses-style endpoints that better support chat, multimodal, and agentic flows. If your clients still call legacy generate-style endpoints, plan refactors to the newer Responses endpoints and updated client libraries.
What you actually need to change
This is not just "switch a key" — it’s an identity redesign: move from static API keys to service account identities and short-lived credentials. Practical migration checklist:
- Replace static API keys with service accounts (or Workload Identity Federation for non‑GCP CI). Create minimal-privilege service accounts and grant only the model-access role(s) required.
- Use short-lived credentials: mint tokens via IAM, Workload Identity, or the metadata service instead of baking secrets into images. Tighten rotation and revocation workflows.
- Audit and update every CI/CD pipeline, Cloud Run service, Cloud Function, and build agent that currently stores standard API keys in secrets managers or environment variables.
If you run many serverless or ephemeral workloads, adopt Workload Identity (Kubernetes KSA->GSA) or federated identities now. Replacing long-lived secrets with identity-bound flows prevents the next embarrassing incident.
Secondary changes you should know about
Google Cloud also rolled out platform changes that will affect deployment and observability patterns.
-
gateway.dev hostnames for model-routing gateways: new Cloud Run–style AI gateways may default to hostnames like GATEWAY_ID-PROJECT_NUMBER.REGION.gateway.dev instead of run.app. That alters DNS, TLS, and ingress automation assumptions — wildcard run.app certificates, ACME automation, and host-based routing rules may need updates.
-
App Topology API billing is shifting to usage-based pricing with a daily free allotment starting in mid‑September 2026. Large microservice fleets that poll topology frequently will see costs where they previously had only operational metrics, which incentivizes sampling, smarter collection windows, and caching of topology results.
Agent sandboxes and policy
The Gemini enterprise agent platform has introduced GA sandboxes for code execution and shell access, plus context-aware access controls and persistent custom instructions in Workspace. Hardened execution environments are necessary when agents interact with internal systems, but they expand your attack surface and create new trust boundaries. Make sure your access controls, audit logging, and egress policies cover these sandboxes so you can trace which agent identity touched what system.
Final take
This cluster of changes signals a broader shift: Google is forcing identity-first practices into AI integration patterns and treating model access as a first-class resource to secure and bill. If your platform still treats Gemini calls like throwaway outgoing HTTP requests anchored by a global API key, consider this your wake-up call. Teams that migrate to scoped service accounts, short-lived tokens, and careful topology sampling will come out safer and cheaper. The ones that don't will face outages, compromised keys, and surprise bills. Remember: credentials are now an architectural decision, not an ops detail.