If your business uses ChatGPT, Claude, or Gemini behind the scenes, you have an AI API key somewhere. Maybe it powers your website chatbot. Maybe it writes product descriptions in your Shopify store or runs an automation your agency set up last year. Last month, thieves proved those keys are worth real money.
Hackers broke into Railway, a popular cloud hosting platform, and ran up an estimated $740,000 in AI usage with API keys stolen from its customers. Railway’s incident report, published November 10, 2026, says the attacker was specifically hunting for AI provider credentials. Here is what happened and the checklist we are running for our clients this week.
What Happened at Railway
The intrusion started on October 20, 2026 through a flaw in Railway’s build systems. Railway spotted suspicious activity the same afternoon and had the attackers fully removed by the morning of October 21.
In that window, the attacker decrypted stored settings for at most 1.85% of Railway workspaces, which works out to roughly 100,000 workspaces. Database logs show they searched for variables named like AI keys, plus cloud, database, and payment credentials. Linked account tokens for 1,852 users were also exposed, including GitHub logins and the AI keys people had connected to agents.
Railway was not alone. OpenCode disclosed a similar theft on October 17 affecting 1.4% of its console users, with some people reporting charges above $1,000. Warp and Fireworks AI reported related incidents in October too.
Security people now call this LLMjacking. Thieves steal your AI key, then resell access to your paid account through shady AI routing services. According to research cited by CSO Online, stolen keys were tested within 30 minutes and then flooded with traffic four to seven days later.
Who Pays the Bill
This is the part that should worry every business owner. Refunds depend on which company issued your key, not on whose fault the breach was.
- AWS Bedrock and Google Gemini started refunds under their existing fraud policies.
- OpenAI is reviewing claims case by case through support tickets.
- Anthropic charges are being covered by Railway while talks continue.
- OpenRouter is sticking to its no-refund terms, so Railway is paying those costs.
On October 26, Railway actually told customers it could not refund this spend at all. It reversed that decision two weeks later. None of the big AI providers has promised refunds for future incidents either. Next time, the bill could land on you.
What the Coverage Missed for Small Teams
I read the top stories on this breach from security outlets and they are solid. But almost all of the advice assumes you have a security engineer on payroll.
Most of the business owners we work with do not have that person. They have a Shopify store with five AI apps installed, a WhatsApp bot a freelancer built, and an automation tool nobody has logged into for months. Nobody knows how many keys exist or where they live. That is the real gap, and that is what the steps below fix.
The One Hour AI Key Safety Checklist
You can do most of this yourself in an afternoon. Where you need a developer, the job is small.
Step 1 List every tool that uses an AI key
Open a simple spreadsheet. Write down every chatbot, Shopify or WooCommerce app, automation workflow, and hosted platform that talks to an AI model. Next to each one, note who created the key and which provider issued it.
Step 2 Give every tool its own key
One shared key across five tools means one leak takes down all five, and you cannot tell which tool caused it. Create a separate key, or better a separate project, for each tool. When something goes wrong, you revoke one key instead of breaking your whole operation.
Step 3 Set hard spending caps and early alerts
OpenAI lets you set hard monthly limits per project. Anthropic lets you set your own monthly limit on the billing page below your tier cap. Google offers caps at the billing account level, with per-project caps still experimental. Set alerts at around half your normal spend and send them to a phone someone checks at night. Caps are not always instant, so the alert is your real early warning.
Step 4 Rotate the right way
Create the new key first, put it into the tool, and confirm everything still works. Only then delete the old key at the provider. Removing a key from your app settings does nothing if the key is still active at OpenAI or Google. If you ever stored keys on Railway or OpenCode before late October, rotate them today.
Step 5 Check your bills every week
Spend ten minutes every Monday looking at usage by key. Look for traffic at odd hours, models you never use, or sudden spikes. If you find something, save screenshots of the billing data and open a support ticket right away. Evidence makes refund claims far easier.
A Scenario We See All the Time
Picture a restaurant group with three branches and a busy online ordering site. A freelancer built them an AI chatbot for WhatsApp orders and put the OpenAI key into a hosting platform. The same key also runs their menu description generator and a review reply tool.
The hosting platform gets breached. Nobody set a spending cap, and the billing alerts go to the freelancer’s old email. Within a week, thieves burn through thousands of dollars. Then the provider freezes the account, and the ordering bot goes dark on a Friday night.
With separate keys, a $150 cap on each, and alerts on the owner’s phone, that same breach becomes a small annoyance. One key gets revoked, the bot keeps taking orders, and the owner files a refund ticket on Monday.
Questions to Ask Your Vendors and Developers
Before you paste a key into any hosted tool, ask three simple things. Where is my key stored? Who inside your company can decrypt it? Can I use a key with its own spending limit? If a vendor cannot answer clearly, prefer tools that keep keys on your own systems.
Frequently Asked Questions
How do I know if my AI API key was stolen?
Check your provider’s usage dashboard and filter by key. Unusual spikes, activity at times your business is closed, or models you never use are strong signs. Many owners only find out when a spend alert or a frozen account shows up, which is why alerts matter so much.
Will OpenAI or Google refund charges from a stolen key?
Sometimes. In the Railway case, Google and AWS refunded under their fraud policies and OpenAI reviewed claims case by case. No provider has promised refunds for future incidents, so treat a spending cap as your real protection.
Do Shopify and WooCommerce apps put my AI keys at risk?
Most store apps use their own AI accounts, so your key is not involved. The risk comes from apps, plugins, or custom features where you pasted your own key. Those are the ones to list, cap, and rotate first.
The Takeaway
AI keys are now treated like credit card numbers by thieves, and your provider may not cover the damage. You do not need a security team to protect yourself. You need a list of your keys, one key per tool, hard caps, and alerts someone will actually see.
If you need a hand putting this into practice, whether that is building or fixing a Shopify or WooCommerce store, developing a custom app or chatbot, or getting unstuck with ads and SEO, our team at ConceptRecall (conceptrecall.com) is happy to have an honest conversation about where you are and what to fix first.