Railway estimates $740,000 in unauthorized AI use after key theft

Platform-as-a-service provider Railway says customer AI provider keys stolen during its October security breach were used for an estimated $740,000 in unauthorized usage charges. Railway said it would pay the affected charges itself when an AI provider does not refund them.
The company hosts and runs customer applications and stores secrets, including API keys and database credentials, as environment variables. Its investigation found that an attacker decrypted variables in at most 1.85% of workspaces, an upper-bound estimate of about 100,000 workspaces, and queried for AI provider keys.
Scope of the exposed credentials
Railway said its logs identify the environments exposed, but not the individual variables within them. The attacker also accessed stored histories for 353 builds, including variables and secrets passed to those builds. A separate query matched credentials for cloud, database and payment services.
The incident report says linked-account tokens belonging to 1,852 users were also decrypted, including BYOK agent keys and GitHub OAuth tokens. Railway reset those tokens and notified each affected user. It said it found no login to a customer account using stolen credentials, no evidence of customer database access, and no indication that customer code repositories were downloaded.
Refund routes differ by provider
Railway said AWS Bedrock and Google Gemini have fraud policies under which their account teams have started refunds. OpenAI and SpaceXSI claims are reviewed case by case: customers submit a claim to the provider and provide the case ID to Railway support.
Discussions with Anthropic remain under way, and Railway said it will cover costs arising from breached Anthropic keys in the meantime. OpenRouter maintains a no-refund policy, so Railway said it will cover all breach-related OpenRouter costs. The incident report does not explain how the $740,000 estimate was calculated, the period it covers, or a deadline or cap for claims.
Attack path from build system to secrets
Railway said the intrusion began at 07:28 UTC on October 20, when an attacker exploited a previously unknown flaw to partly escape filesystem isolation around a build and read files on the host virtual machine. A credential used by build servers then provided access to internal systems after Railway's Temporal service accepted it as a trusted client.
The attacker obtained cached code snapshots and a copy of Railway's internal monorepo, then identified a misconfigured OIDC binding in a cloud environment. At 03:26 UTC on October 21, that binding was used to obtain short-lived cloud credentials. The attacker subsequently obtained credentials able to read encrypted production variables from inside the internal network and an SSH key for an orchestration machine.
Between 04:07 and 05:21 UTC on October 21, the attacker retrieved encrypted variables while looking for AI-key names. Railway said the orchestration credentials enabled decryption through its cloud provider's KMS. The attacker was fully removed at 09:24 UTC that day.
Actions for Railway customers
Railway advises customers to revoke and replace every sensitive variable stored on the platform before October 21, using the provider that issued each key. Reporting an unauthorized charge does not revoke a key. It also recommends removing public database network access and reviewing GitHub, Docker Hub and other registry audit logs.
The provider has moved builds to short-lived virtual machines, with credentials restricted to pushing one specific image to its registry, and says every exploited weakness has been closed. For businesses, the immediate implication is to rotate exposed secrets promptly, reconcile AI-provider bills, and treat environment-variable access as a high-priority control in build and deployment systems.

