Railway Breach: $740K in AI Charges on Stolen Keys
Attackers took AI keys stored on Railway and ran up hundreds of thousands in charges. Here is what a small team should rotate and check this week.
By Peter Walters
Railway, the app hosting platform, published its full incident report on November 10, 2026. Attackers who broke in on October 20 and 21 decrypted customer environment variables and used stolen AI provider keys to run up an estimated $740,000 in unauthorized charges. If you or a developer you hired ever stored an OpenAI, Anthropic, Gemini or similar key on Railway, that key should be treated as stolen and replaced.
What actually changed
Railway's incident report says an attacker escaped the isolation around its build servers using a previously unknown flaw. From there they reached an internal credential, then a misconfigured trust setting that let them decrypt customer environment variables. Environment variables are where apps keep secrets like API keys and database passwords.
Railway says up to 1.85% of its workspaces, roughly 100,000, could have had variables decrypted. It calls that an upper bound, because its logs show which environments were touched but not which individual variables were read. The attacker searched mainly for AI provider keys, and a second search also matched cloud, database and payment credentials. Linked-account tokens for 1,852 users, including GitHub tokens, were also exposed and have been reset.
Railway reports no evidence that customer databases were accessed or that customer code repositories were downloaded. The Hacker News summary lists how providers are handling refunds. AWS Bedrock and Google Gemini say their usual fraud-refund policies apply. OpenAI and SpaceXSI review claims case by case. Railway is covering Anthropic and OpenRouter costs for now, and says it will pay any charges a provider will not refund.
Two things are still unclear: how the $740,000 figure was calculated, and whether there is a deadline for filing a claim.
Why it matters if you run a small team
Most coverage treats this as a story about developers. It is also a story about you, because small businesses now own AI keys without thinking of them as keys. A website chatbot, an inbox sorter and a booking assistant all call an AI provider with a secret that works like a credit card number with no limit printed on it.
The attacker wanted those keys because they turn directly into money. Railway believes the keys went to services that resell AI access. Nobody needs to see your customer list for this to hurt. They only need your key and your billing account.
What is overhyped: the idea that this proves AI tools are unsafe. A hosting company was breached. The same thing could happen to any platform that stores your secrets. What is not overhyped: the pattern. Railway's own report notes that the same campaign touched other AI developer platforms in October, so expect more of these.
What I would do with it this week
If anyone built you an AI tool, send them this list today.
- Ask where your AI provider keys are stored and whether any lived on Railway before October 21.
- If yes, rotate them. Create a new key with the provider, update the app, then delete the old key. Reporting a charge to Railway does not cancel the key.
- Look at the last 30 days of usage in each AI provider dashboard for spikes you do not recognize.
- Report any unauthorized charge using the form in Railway's report, and file a claim with the provider as well.
- Check whether your provider lets you set a monthly spending limit, and set one that matches what you really use.
- Check your GitHub account for apps and tokens you do not recognize.
When I built a bookkeeping and client-ops system, I gave its Stripe connection a read-only key on purpose. The system can look at payments but cannot move money, so a stolen key would be an annoying afternoon instead of a disaster. I use the same thinking on chatbots: one key per project, the narrowest permission available, and a spending limit where the provider offers one. A key you can replace in five minutes is a small problem. A key shared across six apps is a weekend.
You do not need to understand how the attack worked. You need to know which secrets you own, where they live, and how fast you can swap them.
Sources
- Incident Report: October 21st Security Attack, Railway
- Railway says stolen customer AI keys, The Hacker News
- Railway breach leads to $740,000 in fraudulent AI charges, BleepingComputer
- Everything That Happened in AI Today, November 11, 2026, The Neuron
I write about this kind of thing every day, and you can get it in the letter.