Railway breach exposed variables in up to 100,000 workspaces
Attackers were inside Railway's systems for more than a day before eviction on October 21, generating about $740,000 in unauthorized AI charges.
Key points
- Railway said at most 1.85% of its workspaces, about 100,000, could have had environment variables decrypted during the October attack.
- Credentials obtained in the breach generated roughly $740,000 in unauthorized AI charges, and Railway said it will cover costs from its own balances if providers refuse to refund them.
- Linked-account tokens for 1,852 users were also exposed, and Railway reset every one of those tokens and notified each affected user.
- Railway said it has closed every attack path, filed a complaint with the FBI's IC3, and hired a forensics firm whose investigation is continuing.
Railway, a cloud hosting platform, said a mid-October security attack exposed environment variables for at most 1.85% of its workspaces, an upper bound covering about 100,000 accounts. The obtained credentials were then used to rack up roughly $740,000 in unauthorized AI charges across multiple providers.
The company published a detailed incident report after engineers spent more than a day evicting the intruder. Railway said the attacker first gained a foothold at 07:28 UTC on October 20 and was fully removed at 09:24 UTC on October 21.
How the attack unfolded
Railway said the attacker began by exploiting a previously unknown flaw in its build infrastructure at 07:28 UTC on October 20. At the time, Railway's build system ran multiple customers' code on shared virtual machines and relied on file-system isolation to keep those workloads separate. The attacker used the flaw to break partly out of that boundary and grab an internal build credential.
Railway engineers did not spot unusual activity until 13:58 UTC on October 20, when alerts flagged abnormal calls to an internal registry service. The team traced those calls back to the compromised credential and assembled an incident response group. Builds were halted at 17:03 UTC on October 20 while engineers patched the isolation flaw.
What the team did not yet know was that the attacker had already pulled a copy of Railway's internal source code, known as the monorepo. With that blueprint in hand, the attacker scanned for more entry points. Early on October 21, the attacker exploited a misconfigured OIDC binding, a trust relationship used to issue short-lived cloud credentials, and used it to reach the decryption keys for stored environment variables.
Using those credentials, the attacker queried Railway's database of encrypted environment variables, prioritizing rows whose names matched AI provider key patterns. Railway then disabled cloud credentials and halted all builds again at 07:22 UTC on October 21. The attacker was fully gone by 09:24 UTC on October 21. Railway summed it up plainly: "This was not a single breach followed by a single fix," the company said, noting that the attacker and its defenders worked in parallel across multiple fronts.
How the railway mid-October breach exposed customer variables
Railway said the exposure ceiling was 1.85% of workspaces, roughly 100,000, that could have had variables decrypted. The company was clear this is an upper bound, not a confirmed count: logs identified affected environments but not which individual variables were actually read.
The attacker also read stored histories from 353 builds between 13:50 and 15:30 UTC on October 20, picking up sealed variables from those builds in the process. That access came through the initial build credential before Railway revoked it.
Beyond environment variables, the breach also reached linked-account tokens for 1,852 users. These included BYOK agent keys, which connect a user's own AI provider account to Railway, and GitHub OAuth tokens, which provide access to code repositories. Railway said it reset all those tokens and notified each of the 1,852 affected users individually. The company said it found no evidence that credentials obtained in the breach were used to log into any Railway account.
Railway also contacted GitHub, GitLab, Anthropic, OpenAI, OpenRouter, SpaceXSI, Google, and Stripe starting at 09:25 UTC on October 21 and asked each service to revoke potentially exposed keys.
Roughly $740,000 in unauthorized AI charges
Railway estimated unauthorized spending on the obtained keys at about $740,000. The company said it is pursuing refunds from the affected providers and will cover any remaining costs from its own balances for providers unwilling to issue a refund.
As of the report's publication, Railway said it would pay Anthropic and OpenRouter costs itself. AWS Bedrock and Google Gemini account teams had begun issuing refunds. OpenAI and SpaceXSI were reviewing claims on a case-by-case basis.
BleepingComputer reported that Railway connected the incident to a broader wave of AI credential thefts that security firm Wiz has called "Black October." Railway said it believes the attacker's goal was to gather AI provider keys for resale or use in model distillation, a technique for training a smaller AI model using outputs from a larger one. Railway said the unauthorized usage it observed ranged from coding agents to AI training pipelines.
What Railway did to close the gaps
Railway said each attack path has been fully closed. Builds now run on dedicated temporary virtual machines, each provisioned with a narrowly scoped credential. The company rebuilt its internal continuous integration system from verified source code and revamped how secrets are stored, so only the dedicated secrets service can obtain the keys needed to decrypt data-encryption keys.
Railway invalidated every user session and token at 20:19 UTC on October 21. It restored builds for all account tiers by 22:49 UTC on October 22, after hardening the infrastructure. The company also reviewed all service-account privileges across its cloud environment.
"Protecting our customers and rebuilding your trust remains our goal," Railway said in the incident report. The company filed a formal complaint with the FBI's IC3 and hired a specialist forensics firm whose investigation is ongoing.
What customers should do now
Railway asked all customers who notice charges they did not authorize to fill out a form linked from the incident report page. The company said it will cover costs that providers refuse to refund, giving affected users a clear path to recovering unauthorized charges.
Railway also said customers can reach out through the support site named in the report with any questions. The company said it will share technical indicators of compromise, details that help other platforms spot related attacks, only after law enforcement gives approval.
What is still not known
The confirmed number of workspaces whose secrets were actually taken remains unknown. Railway's figure of about 100,000 is a maximum, not a verified count, and the gap between that ceiling and the real total has not been disclosed.
Railway gave no timeline for publishing those indicators of compromise, saying only that it depends on sign-off from law enforcement. That information could help other platforms identify whether they were targeted by the same campaign.
The specialist forensics firm Railway hired is still working. Its findings may clarify how widely the obtained credentials were distributed and whether additional credential types were exposed beyond the categories Railway described.
Sources
- BleepingComputerRailway breach leads to $740,000 in fraudulent AI charges
More from the network
Junious Media Network ↗Slow Horses Season 8 Will Be the Show's Last, Olivia Cooke Says
Olivia Cooke says she is filming what she believes is the final season of Slow Horses, though Apple has made no official confirmation.
junidaily.media ↗Glenn Close, Ridley Scott Receive Honorary Oscars at 17th Governors Awards
Five film veterans, including Glenn Close and Ridley Scott, were honored at the Academy's 17th Governors Awards, kicking off the race for the 2027 Oscars.
junidaily.media ↗California Piers Close as Waves Top 50 Feet This Weekend
Surf exceeding 50 feet battered San Mateo County Saturday, closing piers from Ventura to San Diego and triggering an evacuation warning near Capitola.
junidaily.media ↗Amanda Seyfried to Make SNL Hosting Debut on December 5
Amanda Seyfried hosts Saturday Night Live for the first time on Dec. 5, joined by Don Toliver, while her Elizabeth Holmes documentary drives broad buzz.
junidaily.media ↗