€54k spike in 13h from unrestricted Firebase browser key accessing Gemini APIs

TL;DR Highlight

This is a real-world case where an unlimited API key activated for Firebase AI Logic (Gemini API) was exploited in automated attacks, resulting in €54,000 in charges within 13 hours, and Google refused a refund. It serves as a warning about the dangers of exposing API keys on the client side.

Who Should Read

Frontend/full-stack developers using Firebase or Google Cloud to develop or launch features based on the Gemini API. Specifically, small teams or individual developers who include API keys in client code or fail to properly configure budget management.

Core Mechanics

After activating Firebase AI Logic to add AI features based on the Gemini API (generating web snippets from text prompts) to a project originally created a year ago for Firebase Authentication, automated traffic surged immediately upon activation.
The browser API key for the Firebase project was originally created with 'No API restrictions,' and with the addition of the Gemini API, Gemini requests began to flood in without limit using that key. These were automated bot requests unrelated to actual user traffic.
A budget alert of €80 and a cost anomaly notification were set up, but both were delayed by several hours. By the time a response was possible, approximately €28,000 had already been charged, and the delayed cost reporting resulted in a final amount exceeding €54,000.
The Google Cloud support team rejected the request for a charge adjustment, stating that the usage 'occurred from that project' and was therefore considered valid.
Gemini team lead Logan Kilpatrick commented directly, explaining the current countermeasures: Tier 1 users have a default monthly spending limit of $250, and project spending caps can also be set. However, both features have a maximum delay of 10 minutes.
Google stated that it is moving towards preventing the use of Gemini API with unrestricted API keys. It has also changed the default for new users to generate more secure Auth keys.
Prepaid billing is also being introduced, starting with new users in the US and expanding globally, allowing developers to pre-determine and control their spending.
Google had maintained a principle for over 10 years that 'Google API keys are not secrets (they can be public),' but this principle was broken after the release of the Gemini API. However, many developers are unaware of this change and continue to use keys as before.

Evidence

"The structural issues with GCP budget notifications were heavily criticized in the comments. Because billing events pass through a queue/log for aggregation, the notifications themselves can be delayed by hours. Even setting a hard limit operates based on the last known aggregated value, leaving it helpless against sudden spikes. Many agreed with the criticism that 'this structure is designed to protect the company, not the customer.'\n\nAnother developer who experienced similar damage appeared in the comments. After incurring $26,000 in damages, they requested a refund from the Google Cloud support team, which was initially denied but is currently under review, with advice to 'escalate as much as possible.' Another comment shared a case where a Gemini API key quickly generated during a Google public live training session was leaked, resulting in approximately $6,909 in charges.\n\nIt was pointed out that many Gemini API keys are already hardcoded on GitHub. A search for 'gemini \"AIza\"' on GitHub yields numerous results, and developers are analyzing that they are treating Gemini keys in the same way as existing Google Maps/Firebase keys, which have long been considered safe to make public.\n\nInformation was shared about the 'emergency brake' feature in GCP that immediately stops billing. Separating the billing account from the project stops API usage immediately, but there is a risk of deleting associated resources, making it unsuitable for production apps.\n\nA developer using Backblaze B2 mentioned it as an example of a 'working spending limit.' B2 shares their experience that API requests are immediately blocked when a $0 limit is set, exceeding the free tier, emphasizing the need for a fundamental change in cloud service billing design.\n\nQuestions were raised about why scammers target AI API keys. In response to the question of what clear financial gain there is, like mining Bitcoin with EC2 credentials, speculation arose that it may be to use or resell LLM services under someone else's account."

How to Apply

If you plan to integrate the Gemini API with your Firebase project, you must simultaneously set both 'HTTP referrer restrictions' and 'API restrictions (allow only generativelanguage.googleapis.com)' in the API key settings. Without these two restrictions, the key can be used to call all Google services if exposed.
Design your architecture so that Gemini API calls are made only on the server side (not on the client side - browser, app). If client-side calls are unavoidable, apply Firebase App Check to restrict API calls to authenticated apps only.
Set a project spending cap (project spend caps) in Google AI Studio or the GCP Console, and build a pipeline using Pub/Sub + Cloud Function to automatically disconnect the billing account when the limit is exceeded. Simple email notifications are not a practical defense due to delays of several hours.
If you add the Gemini API to a project originally created for other purposes, such as Firebase Authentication, the existing API keys in that project will all have Gemini access. You must thoroughly review the existing key list and remove unnecessary API permissions.

Code Example

snippet

# GCP billing account disconnection method via programming (based on official documentation)
# Caution: Do not use for production apps. Associated resources may be deleted.

# Disconnect billing account using gcloud CLI
gcloud billing projects unlink PROJECT_ID

# Or use Pub/Sub + Cloud Function combination to automatically disconnect upon budget exceedance
# 1. GCP Console > Billing > Budgets & alerts > Create budget
# 2. Set Alert threshold (e.g., 80%)
# 3. Select 'Connect a Pub/Sub topic'
# 4. Deploy the following Cloud Function

import base64
import json
from googleapiclient import discovery

def stop_billing(data, context):
    pubsub_data = base64.b64decode(data['data']).decode('utf-8')
    pubsub_json = json.loads(pubsub_data)
    cost_amount = pubsub_json['costAmount']
    budget_amount = pubsub_json['budgetAmount']
    project_name = pubsub_json['budgetDisplayName']  # Use budget name as project name

    if cost_amount >= budget_amount:
        billing = discovery.build('cloudbilling', 'v1')
        # Disconnect billing account
        billing.projects().updateBillingInfo(
            name=f'projects/{PROJECT_ID}',
            body={'billingAccountName': ''}
        ).execute()
        print(f'Billing disabled for project {PROJECT_ID}')

# API key restriction settings (Google Cloud Console > APIs & Services > Credentials)
# - Application restrictions: HTTP referrers (web sites) -> Allow only your domain
# - API restrictions: Restrict key -> Select only generativelanguage.googleapis.com

Terminology

Firebase AI LogicA Firebase service that makes it easy to integrate Google's AI features, such as the Gemini API, into Firebase projects. Activation allows Gemini requests to be made with existing Firebase keys.

unrestricted API keyAn API key that can be used without restriction on any API or domain. It is often the default state when a key is first created in Google Cloud.

App CheckA security feature provided by Firebase that verifies that only authenticated apps (iOS/Android/Web) can access Firebase resources. It helps block bots and abnormal clients.

spend capA hard limit that automatically stops services when API usage costs exceed a specific amount. Unlike simple notifications, it actually stops requests.

Pub/SubGoogle Cloud's messaging queue service. It is used to build automated pipelines where messages are published upon the occurrence of events such as budget exceedances, and Cloud Functions subscribe to these messages to take automated actions.

billing aggregation delayThe delay between when an actual API call cost is incurred and when it is reflected in the cloud service's billing system. In Google's case, this can be up to several hours, which is the root cause of delayed budget notifications.