# Railway to pay $740,000 for stolen AI keys

> Railway's report puts fraudulent usage of its customers' AI keys at $740,000 and distributes refunds provider by provider.
> By José Maldonado · Cybersecurity · 2026-11-12

On November 10, 2026, Railway put the fraudulent usage an attacker made with AI provider keys stolen from its customers at about $740,000. The figure appears in [Railway's report on the October incident](https://blog.railway.com/p/incident-report-october-security-incident) that the platform published that day, and it comes with a promise unusual in this type of breach: the company will pay out of its own pocket whatever the providers do not refund. It is not an audit or a closed settlement, but a party's estimate that the company does not explain how it calculated.

Railway sells a platform for deploying applications without setting up servers, and on it teams store their environment variables: connection strings, passwords and third-party API keys. Whoever reads those variables does not get one company's data, but the keys to hundreds.

The assault began in the build system, the one that builds applications before publishing them, and that is the point worth examining closely. From there came internal credentials, histories of other builds and keys with which third parties billed Railway customers. The platform maintains that there is no evidence of access to customer databases or of repository downloads, and that what was copied without decryption can no longer be opened.

## What happened between October 20 and 21?

At 06:50 UTC on October 20 the attacker was probing the build infrastructure. At 07:28 they exploited an unknown flaw, partially escaped the file isolation and read files from the host. From there came the credential that pushes images to the internal registry.

The flaw was not the end of the road, but the door. The internal Temporal instance, the system that orchestrates asynchronous tasks, accepted that build credential as a trusted client, and through it the attacker read the histories of other builds, including their variables and secrets. It is the same type of jump described in the analyses of [guest-to-host escapes](https://observatorioblockchain.com/en/cybersecurity/pwnescape-cve-2026-99999-escape-invitado-anfitrion/)a barrier that gives way at one point and exposes everything sharing the same machine.

With that access he also recovered a copy of the internal monorepo, because one of the builds he read corresponded to the support panel that Railway deploys on itself. The report says it plainly: the attacker had a privileged view of how the platform operates and used it to hunt for new routes quickly. The alert was triggered at 13:58 UTC by anomalous calls to the registry API, and the full expulsion came at 09:24 UTC on October 21.

In between, Railway halted builds in two batches, rotated internal credentials and rebuilt tools from validated code. Deployments that were already underway kept running during the outage, although no one could publish changes.

## An unauthorized charge

Money is the part that has moved the most since the preliminary report. The document acknowledges that unauthorized spending with the stolen keys is estimated at $740,000 and that the company is negotiating refunds with providers while advancing the money for those who refuse. The figure has no public breakdown: the report does not say what period it covers or how it was calculated.

There is a change of stance worth noting. On October 26, when customers began asking about the charges, Railway responded that it could not refund spending made with compromised keys and encouraged them to file claims with each provider. The November 10 report reverses that response and turns the platform into a payer of last resort. The company does not explain the shift, although the outcome is more favorable for whoever was left with the bill.

Provider

Policy applied

Who covers it while

AWS Bedrock

Anti-fraud policy, refunds already initiated

AWS

Google Gemini

Anti-fraud policy, refunds already initiated

Google

OpenAI

Case-by-case review with Case ID

OpenAI or Railway

SpaceXSI

Case-by-case review by support

SpaceXSI or Railway

Anthropic

Open negotiation with their team

Railway

OpenRouter

No refund per contract

Railway

The differences between the rows determine how long it takes a customer to recover their money and who bears the risk of a half-settled claim. Moreover, the charges form does not revoke any key, and it is the channel the report indicates for filing a claim.

## The real scope of the event

The real scope is smaller than the headline suggests and larger than is reassuring. The logs allow us to state that, at most, 1.85% of workspaces, around 100,000, could have had variables decrypted. The report insists that this is a ceiling: the logs identify environments, not individual variables, so there is no way to know how many keys were left exposed.

Tokens linked to accounts are counted. There were 1,852 users with BYOK agent keys and GitHub OAuth tokens, all reset and notified. The company says it has not detected logins to Railway accounts with credentials from the incident, nor any evidence of access to customer databases or repository downloads. What was not decrypted can no longer be decrypted, because the credential that allowed it has been deactivated.

The usual reasonable doubt remains: if the attacker copied variables and never opened them, there is no log to prove it. Railway bases that part of its account on the deactivation of the credential and on its own logs.

The report's operational advice is the same one security teams have repeated for years: revoke and rotate any secret stored on the platform before October 21, and do so with the provider that issued the key. An episode on November 10 illustrates this: the [Claude Code credentials leaked in 1,122 public repositories](https://observatorioblockchain.com/en/ai/claude-code-credenciales-filtradas-settings-local-json/) show how quickly a plaintext key goes from an oversight to a billing problem.

## The October wave nobody was watching

Railway was not an isolated case, and its report admits as much when it describes a series of attacks against platforms that store AI keys. Wiz's tally is broader: at least nine AI development platforms acknowledged incidents between October 6 and 30, and the [security firm's analysis](https://www.wiz.io/blog/black-october-for-ai-developer-platforms) groups them under the label LLMJacking, the theft of others' computing capacity to resell it or to distill models.

The pattern repeats in almost every statement: access to credentials, a deliberate search for environment variables and a quick exit through a market that buys others' tokens. The demand for data to train on and the existence of intermediaries that resell access explain why someone spends a week chaining exploits instead of encrypting a disk and demanding a ransom. The loot is spent on tokens, not negotiated.

For the customer of a PaaS platform, the uncomfortable lesson is that the provider's isolation is one more layer, not a guarantee. Hosting had already shown that chain in incidents such as [cPGuard and its update channel](https://observatorioblockchain.com/en/cybersecurity/cpguard-backdoor-canal-actualizacion-hosting/).

## What remains open in the case

There are loose ends the report leaves untied. Railway filed a complaint with the FBI's IC3, hired an external forensic firm and closed sign-ups until November 2, after pausing them on October 22. The company also says it obtained access to Project Glasswing with the approval of the US Treasury, a program it gives no details about and that does not clarify how it affects those claiming for the charges.

There is also no per-customer cap or deadline for filing a claim, and the report does not answer whether the bill must be paid while the claim is under review. As long as that remains the case, the figure of $740,000 works as what it is: a one-sided estimate, unaudited, that could rise if more charges appear or fall if providers accept more refunds.

The ball, in any case, remains in each customer's court. Railway can pay the bill and even refund it, but it cannot rotate on anyone's behalf the keys issued by AWS, Google or OpenAI.

The October bill says less about a specific attacker than about the economy built around AI keys. As long as a stolen token is instant money, every platform that stores others' secrets will be a profitable target, and rotation will remain the task nobody does on the day they should.
