How does FireShield know my bill is spiking?

Google reports each project's spend to FireShield through its budget notifications. The moment a report arrives, FireShield checks it against your thresholds and runs their actions; nobody has to read an email first. How often and how soon Google reports is up to Google.

What does it actually do when a threshold is crossed?

Whatever you set for that threshold. Every budget starts with a Hard Kill at 100% (disconnecting the project's GCP billing account so spend stops), which you can move or remove. On any threshold you can add alerts (email, Slack, Telegram) and, if you want it, a Soft Kill: a signed webhook to your own infrastructure asking it to degrade service.

Do I need to change my application code?

No. You link your project by signing in with Google, which grants FireShield an IAM role on its billing account. Nothing to install. Your code only changes if you choose to use Soft Kills: then you add one endpoint to receive the webhook. FireShield never touches your running services directly.

Can FireShield read my source code or infrastructure secrets?

No. Onboarding is zero-secrets: it uses IAM role grants, not uploaded credentials.

What's the difference between Starter, Pro and Enterprise?

All three protect your projects the same way: unlimited thresholds, email alerts, Soft Kill and Hard Kill. Starter covers one project. Pro covers up to 5 projects and adds Slack and Telegram alerts, per-project spend charts, a spend breakdown by service, and usage spike detection checked every 10 minutes. Enterprise has everything in Pro, with no limit on projects and spike checks every 5 minutes.

Does FireShield work with Firebase projects?

Yes. Firebase runs on Google Cloud, so any Firebase project on the pay-as-you-go Blaze plan is a GCP project FireShield can protect. If you've ever hit a surprise Firebase usage limit or a Blaze plan bill that got out of hand, that's exactly the scenario FireShield exists to stop.

What is usage spike detection?

Google reports spend with a delay of 15 minutes to several hours, so a runaway can grow before any budget alert fires. On Pro and Enterprise, FireShield also reads the usage of each service (Firestore, Cloud Functions, Storage, Realtime Database, Hosting, Gemini) and compares the last hour with the same hour over the last 7 days. When a service jumps past the multiplier you set, say 10 times its normal level, it runs the actions you chose: an alert, a Soft Kill or a Hard Kill. It measures usage, not cost, and you turn it on per project, which grants FireShield read-only access to that project's usage metrics.

← All horror stories

Bill horror story

€54,000 in 13 hours: a Firebase browser key used for Gemini

October 21, 2026 · 3 min read

After enabling Firebase AI Logic, automated traffic used the project's unrestricted browser key. The €80 budget alert fired hours late.

Retold from Google AI Developers Forum, by zanbezi, April 15, 2026. Summary in our own words; every figure is the source's.

What happened

Shortly after turning on Firebase AI Logic in an existing project, the poster saw automated Gemini traffic coming through the browser API key Firebase had created for the project. The key had no API restrictions. In the poster's words: "The traffic was not correlated with our actual users and appeared to be automated."

About €28,000 had built up before they could react, and the total passed €54,000 in around 13 hours. They had an €80 budget alert and a cost anomaly alert; both fired hours late. The thread reached the front page of Hacker News.

How it ended

Google Cloud support called it valid usage and denied a billing adjustment. A Google employee replied in the thread about new safeguards (spend caps and plans to disable unrestricted keys) and asked the poster to get in touch. The thread doesn't show a final resolution.

What would have stopped it

Restricting the Firebase key. Limiting the browser key to the APIs the app uses, or protecting AI Logic with App Check, closes the door this traffic came through.
Not relying on the budget alert. An €80 alert that fires hours late is a notification after the fact (why alerts lag).
Acting on usage within minutes. AI Logic token usage reaches Cloud Monitoring in minutes. A spike rule on it, with a Soft Kill that turns AI features off or a Hard Kill as a last resort, would have stopped the spend in the first hour.