Google Cloud

JSON Schema Safety Harness for Text-to-Image APIs

The article recommends placing a prompt-safety gate before image generation by using a chat model constrained to return a strict JSON Schema with decisions: allow, block, or review. The author demonstrates a Python notebook-to-production path using an OpenAI-compatible interface and highlights trade-offs between providers: use a broad consistent REST surface (e.g., Infrai) when you expect multiple backend capabilities under one contract, but prefer provider-native moderation when minimal preflight latency or independently managed moderation is required. The post includes a runnable example, retry semantics, an evaluation methodology, and metrics teams should measure (unsafe-prompt recall, false-block rate, schema-valid response rate, latency, and token costs).

ZoomInfo Adds Gemini Enterprise Connector for Buyer Intent

ZoomInfo has announced a native connector that integrates its headless GTM context layer with Gemini Enterprise, Google Cloud’s enterprise AI platform. The integration allows Gemini Enterprise agents to query ZoomInfo’s verified account, contact, and buyer intent data in natural language by configuring ZoomInfo as a data store in the Gemini Enterprise console. The connector uses ZoomInfo’s GTM Context Graph (covering more than 100 million companies and 500 million contacts) and enforces governance — access control, permissioning, data lineage, and audit logging — tied to customers’ Gemini Enterprise and ZoomInfo permissions. The connector is in public preview and available to Gemini Enterprise customers across global location options; Google Cloud provides setup documentation for administrators.

Google Ads API pilot secures manager-account API access

Google is piloting a security feature for the Google Ads API that restricts sensitive API methods (account management, user management, billing) for accounts under a Google Ads manager hierarchy to a pre-approved allowlist of Google Cloud projects. Participating developers submit the top-level manager customer ID; Google audits API activity to identify associated applications and project numbers, then works with the manager to establish the allowlist. Unapproved applications attempting sensitive calls will fail. The pilot promises enhanced security and improved visibility into which applications can perform sensitive actions. Developers can apply via a Google form; allowlisting changes and opt-out changes are performed within 10 business days.