Railway, whose platform builds and deploys customer applications, has published its full report on the October attack on its systems. It estimates unauthorized spend on stolen keys at about $740,000 and says the attacker decrypted customer variables for at most 1.85% of workspaces, about 100,000.
According to the report, the attacker partially escaped filesystem isolation in Railway’s multi-tenant build infrastructure at 07:28 UTC on 20 October, took a credential the build servers use, and found that an internal Temporal instance accepted it as a trusted client. Railway says the attacker read the stored histories of 353 builds, moved laterally and was fully evicted at 09:24 UTC on 21 October. Railway’s forensics point to AI provider keys as the main target. It asked GitHub, GitLab, Anthropic, OpenAI, OpenRouter, SpaceXSI, Google and Stripe to revoke potentially exposed keys, and reset linked-account tokens for 1,852 users. Railway calls the 1.85% figure an upper bound, because its logs identify environments, not individual variables.
The cost landed on AI provider bills. Railway says AWS Bedrock and Google have started refunds under existing fraud policies, OpenAI and SpaceXSI review claims case by case, and Railway will cover costs itself where a provider will not, including OpenRouter, and Anthropic in the interim. Railway was one of the platforms in Wiz’s tally of AI developer platform incidents earlier this month.
In our reading, the control that capped loss for affected customers sat at the AI provider: spending limits and usage alerts on each key. Teams that kept provider keys in any hosting platform’s environment variables should confirm every such key has been rotated since 21 October, set hard spend limits per key, and report unrecognised charges through Railway’s claim form. Our earlier argument to treat cloud billing admin as a privileged role applies to AI provider consoles too.
Source: Railway: Incident Report: October 21st Security Attack, 10 November 2026