Run · founder · 6 min read
Railway's Final Breach Report: $740,000 in Stolen AI Key Spend, and What to Do
Railway's Nov 10 report says attackers decrypted customer secrets and ran up about $740,000 on stolen AI keys. Who pays, and what to rotate.
On November 10, Railway published its full incident report on the attack it first disclosed on October 21. The short version: an attacker got into Railway’s build systems, reached the database where customer secrets are stored, decrypted some of them, and went straight for AI provider keys. Railway estimates the unauthorized spend on those keys at about $740,000.
Railway is the back-end host we recommend most often for AI-built apps, so this one lands close to home. If you had a project there before October 21, there is a short list of things to do, and a form to fill in if someone ran up a bill on your key.
What happened
All of this comes from Railway’s own three posts: the October 21 disclosure, the October 22 preliminary report and the November 10 final report.
- October 20: the attacker exploited a flaw in Railway’s shared build infrastructure and partly escaped the isolation between customers. That exposed a build-server credential, which let them read the stored histories of 353 other customers’ builds, secrets included. Railway’s monitoring paged its on-call team that afternoon and builds were halted.
- October 21: a misconfigured cloud permission gave the attacker a short-lived credential. From there they reached a machine holding a long-lived decryption key, and used it to decrypt a subset of customer environment variables. Railway says the attacker was fully evicted by 09:25 UTC, and all user sessions and tokens were invalidated that evening.
- October 22: builds came back, enterprise customers first. New signups stayed closed until November 2.
Railway also says the attacker took a copy of Railway’s own internal code along the way.
What was exposed
Environment variables are the settings where your app keeps its secrets: the OpenAI key, the Stripe key, the database password. Railway stores them encrypted. The attacker got the means to decrypt them.
Here is the scope, in Railway’s words and numbers:
- Up to about 100,000 workspaces, or at most 1.85% of the total, could have had variables decrypted. That is a ceiling, not a count of victims. Railway’s logs show which environments were touched, not which individual variables.
- AI provider keys were the priority. The attacker’s queries went after those first. A further query matched cloud, database and payment service credentials.
- 1,852 users had linked-account tokens exposed, meaning GitHub connections and bring-your-own-key agent keys. Railway reset all of them.
- Databases: Railway says its forensics did not indicate customer databases were accessed.
- Your source code: the October 22 report said Railway wasn’t ruling it out. The final report says there is no indication the attacker downloaded customer repositories or snapshots.
Nobody logged into a Railway account using stolen credentials, according to the report. The investigation is still open, with an outside forensics firm engaged and a complaint filed with the FBI.
The $740,000, and who pays it
Stolen AI keys are cash. Someone else runs their workload on your account and you get the invoice. Railway’s estimate is that unauthorized spend on the stolen keys reached roughly $740,000. It says the keys were used through reseller services, and it hasn’t established whether the original attacker ran those or passed the keys on.
The important change is on refunds. On October 22, Railway’s FAQ said it could not refund customers for spend on compromised keys. The November 10 report reverses that. Railway now says it is working with providers on refunds and will cover the cost from its own balance where a provider won’t.
By provider, per the final report:
- AWS Bedrock and Google Gemini: refunding under their existing fraud policies.
- OpenAI: case by case, with Railway’s support team acting as the bridge.
- Anthropic: Railway says it will cover the breached-key costs itself while talks continue.
- OpenRouter: no refunds from the vendor, so Railway is covering those costs too.
That is a better outcome than most breaches produce. It is also not automatic. You have to claim it.
What to do this week
1. Rotate everything that lived on Railway before October 21
Railway’s recommendation covers every customer who signed up before that date. Don’t try to work out whether you were in the 1.85%. Rotate all of it: AI keys, payment keys, database passwords, private registry credentials.
Remember the order. Delete the old key at the provider first, then create the new one. Making a new key does not switch the old one off.
2. Check your AI bills from October 20 onward
Open the usage page at each AI provider and look for spend you don’t recognize. If you find some, report it through the form linked in the “Unauthorized token spend” section of Railway’s final report, or contact Railway support at station.railway.com. Do this before you forget which charges were yours.
3. Close your database to the public internet
Railway specifically recommends removing public network access from databases. If your app and database both live on Railway, they can talk over the private network and the public address isn’t needed. If you aren’t sure, this is a ten-minute job for a developer.
4. Review GitHub
Check your GitHub audit log and the list of apps with access to your repositories. Remove anything you don’t recognize. If you pull images from a private registry such as Docker Hub, check that audit log too.
Should you leave Railway?
We aren’t telling you to. Compare this with Zeabur, which was breached twice in two months through basic housekeeping failures and is shutting down. Railway’s secrets were encrypted, the attack took a chain of several flaws, the company published three reports with a minute-by-minute timeline, and it is paying for losses it initially said it couldn’t cover.
It has also changed how the platform works. Every build now runs in its own short-lived virtual machine instead of on shared servers. Every encrypt and decrypt operation is logged centrally. Internal credentials are short-lived wherever possible.
That said, the uncomfortable facts stand. A long-lived decryption key was sitting on a machine an attacker could reach, and the security disclosure telling customers to rotate came a day after the first intrusion. If you hold regulated or highly sensitive data, ask Railway for the details before you renew.
The lesson for any host
Your host holds your keys, and encryption only helps until someone steals the thing that decrypts them. Two habits limit the damage wherever you deploy:
- Set a spending cap on every AI key. Most providers let you cap monthly spend per key or per project. A capped key turns a five-figure surprise into a small one.
- One key per app. Shared keys mean one breach forces you to touch everything.
We cover where secrets leak on your own machine in our guide to .env files and coding agents.
Sources reviewed: Railway’s security disclosure (October 21), preliminary report (October 22) and final incident report (November 10, 2026).
Related guides
founder · 8 min read
Researchers put a virus in an agent's memory file. It spread to the next agent.
New Anthropic/EPFL research says AI agents can pass self-replicating instructions through files like CLAUDE.md. The fix is one paragraph.
founder · 9 min read
Your Agent Can Now Spend Your Money. Nobody Can Prove You Said Yes.
AWS shipped autonomous agent payments this month. The infrastructure works. The paper trail that would settle a disputed charge doesn't exist yet.
founder · 8 min read
35 Security Holes in One Month: Why Vibe-Coded Apps Are Getting Riskier in 2026
35 new CVEs in March 2026 were traced to AI-generated code. Here's what happened and what founders need to do about it.
Enjoying this guide?
Get weekly practical guides, plus tool updates and implementation playbooks.