How API Abuse Nearly Bankrupted Our Startup

On Monday morning, I opened an email from OpenAI with the subject line "Usage Alert - Action Required." The number inside made me physically ill.

The Horrifying Discovery

Our typical OpenAI bill was $500/month. This weekend had generated nearly 100x that amount. Someone had found our AI endpoint and used it as their personal GPT-4 server.

How They Found Us

We built an AI writing assistant. The frontend sent requests to /api/generate which proxied to OpenAI. Simple, clean, and completely unprotected.

// Our original (terrible) endpoint
app.post('/api/generate', async (req, res) => {
  const { prompt } = req.body;

  // No auth check. No rate limit. Just vibes.
  const completion = await openai.chat.completions.create({
    model: "gpt-4",
    messages: [{ role: "user", content: prompt }]
  });

  res.json(completion);
});

The attacker probably found us through:

• Scanning JavaScript bundles for API endpoints
• Checking network requests in browser dev tools
• Automated scanners looking for unprotected AI endpoints

Once they found a working endpoint, they set up a script to pump through requests 24/7.

The Recovery

I immediately rotated the API key to stop the bleeding. Then began the painful process of dealing with the bill.

OpenAI support was understanding but needed assurance this wouldn't happen again. We documented everything:

• Proof that requests came from unknown IPs, not our users
• Our new authentication and rate limiting implementation
• Spending alerts configured at multiple thresholds
• Monthly spending caps enabled

After two weeks of back-and-forth, they forgave $42,000 of the charges. We still paid $5,000, which was painful but survivable.

The Fix

// The secure version
import { rateLimit } from 'express-rate-limit';
import { authMiddleware } from './auth';

const aiLimiter = rateLimit({
  windowMs: 60 * 1000, // 1 minute
  max: 10, // 10 requests per minute per user
  message: 'Too many requests, please slow down'
});

app.post('/api/generate',
  authMiddleware,        // Require login
  aiLimiter,             // Rate limit
  usageTracker,          // Track per-user usage
  async (req, res) => {
    const user = req.user;

    // Check user's remaining quota
    if (user.aiCredits <= 0) {
      return res.status(403).json({
        error: 'Usage limit reached'
      });
    }

    // Process request...
    await decrementUserCredits(user.id);
});

We also set up proper OpenAI spending controls:

• Hard spending cap: $1,000/month maximum
• Email alerts: At $100, $300, $500, $800
• Daily monitoring: Automated checks for unusual spikes
• Per-user limits: Maximum requests per user per day

::

::