/

Railway Says Stolen AI Keys Ran Up About $740,000. It Will Cover Gaps

Railway says October attackers stole customer AI keys and ran up an estimated $740,000 in usage. It pledges to cover what providers will not refund. Exposure was at most about 100,000 workspaces.

By

·

Written with AI assistanceHow we reportCorrectionsReport an error

Rows of black server cabinets with mesh doors stand in a brightly lit white data-center aisle

Railway, the cloud platform developers use to deploy apps, says attackers who broke into its systems in October stole customer AI provider keys and ran up an estimated $740,000 in unauthorized usage. In a Nov. 10 incident report, last updated the same day, the company said it is working with those providers on refunds and will cover charges that the vendors will not repay. The figure is Railway’s estimate. The report does not say how it was calculated or over what period the charges landed.

Key points

  • Railway says attackers were inside its systems from 07:28 UTC on Oct. 20 until they were fully evicted at 09:24 UTC on Oct. 21, after a then-unknown flaw in its multi-tenant build infrastructure.
  • The company says at most 1.85% of workspaces, about 100,000, could have had environment variables decrypted, an upper bound, and that unauthorized spend on stolen keys was estimated at about $740,000.
  • Linked-account tokens for 1,852 users, including GitHub OAuth tokens and bring-your-own-key agent keys, were reset, Railway says, and it found no logins to Railway accounts with stolen credentials.
  • Railway says it will cover unauthorized token costs that providers will not refund, including Anthropic charges in the interim and all OpenRouter costs tied to the breach.
Rows of black server cabinets with mesh doors stand in a brightly lit white data-center aisle
A server room at GIP Networks, photographed April 19, 2017. This is not Railway’s infrastructure and does not show the systems in the October incident. Photo by Acirmandello via Wikimedia Commons, CC BY-SA 4.0. Cropped and resized for this post.

What Railway says was taken

The report, signed by Jacob Cooper, Angelo Saraceno, Sarah Bedell and Luca Casonato, is a follow-up to a preliminary write-up issued while recovery was still underway. Engineers first became aware of unauthorized access at 13:58 UTC on Oct. 20. The company now dates the initial exploit to 07:28 UTC that morning and the attacker’s full eviction to 09:24 UTC the next day, a stretch of roughly 26 hours. Dummy accounts, all unsuccessful, had started two days earlier, on Oct. 18, the company says.

Railway stores app secrets, including API keys, as environment variables. Its report describes a design in which each environment’s variables are encrypted with a per-environment key, itself wrapped by a root key that sits in an upstream cloud provider’s key-management service. From logs, Railway says it can only show that at most 1.85% of workspaces, about 100,000, “could have had any of their environment variables decrypted.” The logs identify a subset of environments, not individual variables, “so that figure is an upper bound.”

The decrypted data, Railway says, also included linked-account tokens for 1,852 users: bring-your-own-key agent keys and GitHub OAuth tokens. The company says it reset all of those, notified each user, and “identified no logins to Railway accounts using credentials taken in this incident.” The attacker also read the stored histories of 353 builds between 13:50 and 15:30 UTC on Oct. 20, “sealed variables included.”

An additional database query “also matched cloud, database and payment service credentials.” Railway says forensics “did not indicate customer databases were accessed,” and that it has “no indication that the attacker downloaded snapshots or repositories of Railway customers.” Those are the company’s forensic conclusions. A specialist firm is still investigating, and Railway says it has filed a complaint with the FBI’s Internet Crime Complaint Center.

The keys were the point

Railway’s forensics, it says, show the attacker was after AI provider credentials, “that they could sell to distributed transfer stations for use in distillation.” The company says it reached that conclusion by reading the attacker’s database queries, which selected rows whose variable names matched expressions for AI keys, and which also “specifically target[ed] customers with large numbers of variables in their environment.” Starting at 09:25 UTC on Oct. 21, Railway notified GitHub, GitLab, Anthropic, OpenAI, OpenRouter, SpaceXSI, Google and Stripe and asked them to revoke potentially compromised keys, providing hashes of every key it held because it could not yet say which ones had been exposed.

Unauthorized spend on the stolen keys “was estimated to reach ~$740,000,” the report says. Railway is coordinating refunds and “backstopping it from our own balances for those unwilling to do so.” As of the Nov. 10 writing, it described vendor policies like this:

  • AWS Bedrock and Google Gemini: existing fraud-refund policies, with account teams already initiating refunds, Railway says.
  • OpenAI and SpaceXSI: case-by-case reviews, with customers told to open a claim and route it through Railway support.
  • Anthropic: talks were still underway; Railway said it would cover breached Anthropic key costs in the interim.
  • OpenRouter: the vendor is enforcing a no-refund policy in its terms, so Railway said it will cover all OpenRouter costs associated with the breach.

Customers are asked to report charges they did not make through a company form. The report still does not cap claims, set a deadline, or explain the $740,000 estimate. Those gaps matter if you are waiting on a credit.

How far the intrusion went, in Railway’s telling

The company describes a then-unknown vulnerability in its multi-tenant build infrastructure, where more than one customer’s build could run on the same virtual machine, that let the attacker read some files on the host, including a credential build servers used to talk to internal systems. From there, the report says, the attacker moved through Railway’s own systems, obtained a copy of Railway’s internal monorepo, and later used internal credentials to pull encrypted variable envelopes from a production database. We are not going to walk through that path. The useful fact for customers is the outcome Railway now admits: some of their secrets were in reach, and AI keys were at the front of the line.

Railway halted builds twice, at 17:03 UTC on Oct. 20 and at 07:22 UTC on Oct. 21, rotated internal credentials, and invalidated every user session and token at 20:19 UTC on Oct. 21. It says existing deployments kept running, though revoked keys caused some services to break. Builds came back in stages on Oct. 22, starting with Enterprise workspaces at 03:32 UTC and ending with Free and Trial at 22:49 UTC, after what the company calls a rebuilt, more isolated build system. New signups were paused from Oct. 22 at 19:06 UTC until Nov. 2.

The report also leans on an unusual detail for a hosting post-mortem: Railway, citing Wiz, describes the fight as “a contest of agent speed,” with the company “defending against abliterated models and a harness similar to ARTEX.” That is a vendor characterization, not something we independently verified. It does, however, fit the larger pattern of this story. The thing stolen was not a database of credit cards. It was the keys that let someone else run expensive models, cheaply, in someone else’s name.

The catch, and what to watch

Our read: the 100,000-workspace figure is an upper bound, and Railway is clear about that. The $740,000 figure is an estimate with no public worksheet. Both can be true at once, and both can still move as the outside forensics firm finishes its work. The more durable point is what the attacker went looking for. Environment variables have always been a tempting pile of secrets. In 2026, the most valuable names in that pile are often AI keys, because they can be resold for distillation or burned as unattended inference. A host that holds those keys is now a distillation-supply warehouse, whether it asked to be one or not.

If you run workloads on Railway, the actionable items are the boring ones the report already implies: rotate any key that ever lived in a Railway environment during the incident window, not only the ones the company emailed about; check provider bills for usage you did not run; and file the charge form if something looks wrong. Railway says IOCs will stay private until law enforcement signs off.

What to watch next is whether the $740,000 estimate holds, whether Anthropic’s refund talks produce a policy or leave Railway on the hook, and whether the outside forensics report revises that 100,000-workspace ceiling. Also watch whether other developer platforms start treating AI keys in environment variables as a special class of secret, with tighter isolation than a shared build box can give them.

Hype in. Signal out.

Lead image: A server room at GIP Networks, photographed April 19, 2017. It is not Railway’s infrastructure and does not show the systems in this incident. Photo by Acirmandello via Wikimedia Commons, CC BY-SA 4.0. Cropped and resized for this post.