Railway has published the full incident report on the October attack it first disclosed while still recovering, and it reframes the incident: the intruders were not chasing source code or payment data first — they were hunting AI provider API keys to feed model distillation.
Published on November 10 by Railway engineers Jacob Cooper, Angelo Saraceno, Sarah Bedell and Luca Casonato, the report traces a roughly 26-hour intrusion. The attacker began probing the platform's build infrastructure on October 20 and exploited a previously unknown flaw to partially escape filesystem isolation. From there it chained a build credential that Railway's internal Temporal orchestration service wrongly accepted as a trusted client, a misconfigured OIDC binding in a cloud environment, and a root decryption credential sitting on an orchestration machine. That path reached the production database of envelope-encrypted environment variables.
Railway says at most 1.85% of workspaces — about 100,000 — could have had variables decrypted, an upper bound. Database logs show the attacker prioritised rows whose variable names matched AI provider key patterns and favoured accounts with large numbers of variables. The actor also read the stored histories of 353 builds, sealed variables included, and obtained a copy of Railway's internal monorepo; Railway says that let it probe for further entry points at speed, describing the escalating fight as a contest of agent speed against abliterated models used without safeguards.
Unauthorized spend on the stolen keys is estimated at roughly $740,000, and linked-account tokens — BYOK agent keys and GitHub OAuth tokens — for 1,852 users were decrypted. Running services stayed online, but builds and deployments were paused for days and new signups were closed until November 2. GitHub, GitLab, Anthropic, OpenAI, OpenRouter, SpaceXSI, Google and Stripe were asked to revoke keys; Railway says it will cover losses where providers will not, including Anthropic and OpenRouter charges, and has reported the attack to the FBI.
The strategy point matters most. Railway links its case to thefts at OpenCode, Warp and Fireworks AI between October 6 and 23, and to Wiz's Black October research documenting at least nine AI developer platforms breached that month with the same goal. Stored AI credentials have become a resellable commodity — and a target worth automating against.
Sources
- blog.railway.comIncident Report: October 21st Security Attack — Railway Blog
- blog.railway.comPreliminary Report: Oct 21st Security Incident — Railway Blog
- thehackernews.comRailway Says Stolen Customer AI Keys Led to $740,000 in Estimated Unauthorized Usage Charges — The Hacker News
- wiz.ioBlack October for AI Developer Platforms — Wiz Research

