Enterprise AI projects can create a costly security problem when prototype credentials survive into production. Google Cloud, part of Alphabet Inc. (NASDAQ: GOOGL), is advising teams to replace raw API keys with managed identities, a shift intended to make model access more traceable and reduce the exposure of credentials embedded in code, developer environments or application configurations.

The guidance focuses on teams that begin with a key from Google AI Studio and later add users, services and production workloads without changing the authentication model. Google said a raw key acts as a bearer token: whoever has it can use it, while the key is not inherently tied to an individual, machine or workload. If such a credential reaches a public repository, chat channel or mobile application bundle, it can be used to run workloads until the organization detects and revokes it.

For developers, Google recommends Application Default Credentials rather than shared keys. Each person signs in with a work account, and the Google Cloud SDK stores a short-lived credential that supported tools can use. That approach replaces the practice of copying secrets between shell profiles, environment files and project settings with an identity that administrators can grant or remove through existing access controls.

The company makes the same separation for applications. A production workload should run under its own service account, Google said, whether it is deployed on Cloud Run, Google Kubernetes Engine or Compute Engine. The environment can supply a short-lived token for that service account, refreshed automatically, rather than requiring a long-lived key in the application code. The documented model gives a team separate identities for developers and production software, so a development mistake does not automatically carry the same permissions or spending exposure into a customer-facing service.

Identity alone does not set an appropriate authorization boundary. Google recommends granting the narrowest role needed to call models, identifying the Vertex AI user role as one example, and using a group rather than separate individual permissions for employees. The company also points to organization policies that can prevent creation of new API keys and service-account keys, while model-access policies can restrict which models a project may call.

Some credentials cannot be eliminated because applications use third-party services. Google’s recommendation is to place those remaining values in Secret Manager and grant access to the relevant identity rather than retain them in a repository or shared file. The approach also makes access events attributable to a person or service account and allows a team to rotate a secret version without copying the value among developer machines.

The post extends the control model to cost and runtime protections. Google described spend-cap budgets, currently in preview for specified services, as a way to pause new requests after estimated spending reaches a set amount. It also highlighted quota limits, monitoring and prompt-and-response screening through Model Armor. Those features address different risks: identity limits who can call a service, while budgets and monitoring help constrain or identify unexpected usage after access has been granted.

For CIOs and platform leaders, the immediate benefit is clearer accountability over who or what can invoke a production model. The guidance does not make credentials risk-free; teams still need inventories, role reviews and incident processes. But it offers a practical path away from unauditable shared keys as AI pilots gain users and spending. Legacy prototype practices treat a secret as a convenient application setting. Google Cloud’s identity-first design treats model access as a governed enterprise service, raising the expectation that AI platforms must connect authentication, authorization, cost controls and auditability rather than leave those functions to scattered developer configuration.