# Railway's breach report: stolen AI keys ran up ~$740,000, up to 100,000 workspaces exposed. Who pays, and what to rotate

> Railway's full post-mortem says an attacker escaped its build sandbox, decrypted customer environment variables and used stolen AI keys for an estimated $740,000 of usage. Railway now says it will cover charges providers won't refund, reversing its earlier stance. Here's what happened and what to do if you deploy on it.

Published Nov 11, 2026 by TechMeasured Staff · Security · Canonical: https://techmeasured.com/security/railway-breach-report-740000-stolen-ai-keys-100000-workspaces-who-pays-what-to-rotate

App-hosting platform **Railway** has published its full incident report on the October intrusion, and the headline number is new: an attacker used AI provider keys taken from customers' environment variables to run up an estimated **$740,000** in unauthorised usage. The November 10 report also says the intruder could have decrypted variables in **at most 1.85% of workspaces — about 100,000** — and took linked-account tokens for **1,852 users**.

Why it matters: a hosting platform has to decrypt your secrets to inject them at runtime, so one breach there exposes every key you stored, not just the ones for that app. And Railway has changed its position on who pays. Its October preliminary report said it was "unable to refund customers" for spend on compromised keys; the final report says Railway will backstop losses that AI providers won't refund.

## How the attacker got in

Railway's report describes a chain, not a single bug (all times UTC):

| When | What Railway says happened |
|---|---|
| Oct 18 | Attacker began creating dummy accounts; early probes failed |
| Oct 20, 07:28 | A previously unknown flaw let the attacker partly escape filesystem isolation on a shared build host and take a build-server credential |
| Oct 20, 13:50–15:30 | Railway's internal Temporal task system accepted that credential as trusted; the attacker read the histories of 353 builds, including sealed variables, and pulled a copy of Railway's monorepo |
| Oct 20, 13:58 | An alert on anomalous registry API calls paged on-call engineers |
| Oct 21, 03:26 | Using the stolen code, the attacker found a misconfigured OIDC binding and got short-lived cloud credentials, then SSH'd into an orchestration machine holding a long-lived decryption credential |
| Oct 21, 04:07–07:15 | Encrypted variables were queried and a subset decrypted via the key management service |
| Oct 21, 07:14–09:25 | Railway disabled cloud credentials and the decryption service; it says the attacker was evicted and had no customer-data access after 09:25 |

That is roughly 26 hours from first exploit to eviction. Railway says the attacker prioritised variables whose names matched AI provider keys and customers with many stored variables, and that it faced "abliterated" AI models — models with safety restrictions removed — driving the attack at speed.

Railway says it found no evidence the attacker downloaded customer repositories or accessed customer databases, and no logins using stolen credentials. The 1.85% figure is an upper bound: Railway's logs identify environments touched, not individual variables.

## Who refunds the fraudulent AI usage

Railway's report lists provider positions "as of writing":

| Provider | Position, per Railway |
|---|---|
| AWS Bedrock, Google Gemini | Existing fraud policies apply; refunds started |
| OpenAI, SpaceXSI | Case-by-case review |
| Anthropic | Discussions ongoing; Railway covering breached-key costs meanwhile |
| OpenRouter | Enforcing its no-refund terms; Railway says it will cover those costs |

Railway says logs from customers show stolen keys were later used for coding agents and AI training pipelines. It suspects resale or model distillation but says it hasn't established whether the attacker ran those resale services. iTnews places the incident in a wider wave: it reports Wiz counted at least nine AI developer platforms disclosing incidents in October, a list that has since grown to 15.

## What to do if you use Railway

1. **Rotate at the issuer, not just in Railway.** Create new AI, cloud, registry, payment and database credentials at each provider, update Railway, redeploy, then revoke the old ones. Railway's guidance treats anything stored before October 21 as potentially exposed. Include sealed variables: Railway says the build histories the attacker read contained them.
2. **Check AI provider usage and billing** for October 20 onward. Report unauthorised charges through the form in Railway's incident report so it can pursue refunds or cover them.
3. **Check cloud accounts** whose keys lived in Railway for new users, access keys, instances or roles, and unusual regions.
4. **Reconnect GitHub and review its audit log** and installed OAuth apps; Railway reset all exposed tokens but recommends the review. Recreate Railway CLI and CI tokens, which were invalidated on October 21.
5. **Redeploy your latest commit** — Railway says commits pushed while builds were paused weren't deployed.
6. **Close public database access** where you can, and review service logs for unexpected connections.
7. **Shrink the blast radius next time:** separate AI keys per service, hard spend caps and alerts on every LLM account, and short-lived credentials (OIDC/workload identity) instead of static keys where supported.

## What Railway changed

Railway says every build now runs on its own ephemeral VM with a credential that can only push one image, the shared build systems that were exploited are retired, CI has been rebuilt with 2FA and biometric approval for critical actions, and all encrypt/decrypt operations are centrally logged. It has filed a complaint with the FBI's IC3, and an outside forensics firm is still investigating, so some details may change.

**Who should care:** anyone who deployed on Railway before October 21 — and anyone storing high-value AI keys on any PaaS. If your platform holds keys that can spend money, assume a breach there means a bill for you.


## Sources

1. [Railway — Incident Report: October 21st Security Attack](https://blog.railway.com/p/incident-report-october-security-incident)
2. [Railway — Preliminary Report: Oct 21st Security Incident](https://blog.railway.com/p/preliminary-report-october-security-incident)
3. [iTnews — App hoster Railway attackers ran up US$740,000 on stolen AI keys](https://www.itnews.com.au/news/app-hoster-railway-attackers-ran-up-us740000-on-stolen-ai-keys-630277)
4. [WindowsForum — Railway breach exposes developer secrets: up to 100,000 workspaces at risk](https://windowsforum.com/news/railway-breach-exposes-developer-secrets-up-to-100-000-workspaces-at-risk-and-740-000-in-ai-usage.451735/)
