Log date · Tue
Android's cloud and hybrid inference guide: what App Check covers and what it does not
The Android blog showed Firebase AI Logic and hybrid inference with App Check against API abuse, while prompts still lived in client code.
What happened on 21 July
On 21 July 2026 the Android Developers Blog published its Build intelligent Android apps series, including a post on cloud and hybrid inference. It showed a sample app calling cloud Gemini models through Firebase AI Logic with the Google AI backend, and switching between on-device Gemini Nano and the cloud through the Firebase hybrid inference API, with modes PREFER_ON_DEVICE, PREFER_IN_CLOUD, ONLY_ON_DEVICE and ONLY_IN_CLOUD. The post named API key abuse and unauthorized billing as risks of cloud AI calls and positioned App Check as the answer, using Play Integrity in production and the Debug provider for local development. The code shown installs only the Debug provider, and system instructions and prompts are defined in client code.
What it means for your threat model
Adding AI features introduces two new assets to protect: the cost and quota of cloud model calls, and the prompts, system instructions and routing logic themselves. App Check targets the first. It helps the backend judge whether a request comes from your genuine, unmodified app, which raises the cost of bulk abuse, but it is not a confidentiality mechanism. System instructions, personas and routing rules compiled into the APK can be read like any other client code. If those prompts carry your product differentiation, assume a competitor can study them directly; if you need them unseen, they have to run on the server.
Checks before shipping
First, confirm that release builds install the Play Integrity provider rather than the Debug provider, and turn on enforcement for the relevant APIs in the Firebase console; keep debug tokens only on developers' own allow lists. Second, inventory the prompts and system instructions in the client, and move anything that embodies core value or contains internal information into server-side templates, with the client sending only the parameters it must. Third, set per-user and per-device quotas and alerts for cloud calls, treat App Check as one layer among several, and log unusual call patterns on the server.
Your checklist
- Confirm release builds use the Play Integrity provider with enforcement on
- Move core prompts and system instructions to the server
- Set quotas, alerts and anomaly logging for cloud AI calls
Primary sources
- Android Developers Blog: Build intelligent Android apps: Cloud and hybrid inference · Source date 2026-07-21 · Checked 2026-10-12
- Android Developers Blog: Build intelligent Android apps: On-device inference · Source date 2026-07-21 · Checked 2026-10-12