Stolen AI keys taken in the October breach of cloud hosting platform Railway ran up an estimated $740,000 in unauthorised AI usage, and whether a customer gets that money back depends on which AI provider issued the key. Railway’s incident report, published on November 10, 2026, lists six providers with four different answers, from automatic fraud refunds to a flat refusal. Railway now says it will pay whatever the providers will not. For engineering leaders, the report is a clear record of something easy to overlook: an AI provider key works like a payment card, and every one a team holds comes with its own issuer’s fraud policy.
Railway Says the Attacker Went Looking for AI Provider Keys
The intrusion ran from 07:28 UTC on October 20 until the attacker was evicted at 09:24 UTC on October 21, according to Railway. It began with a previously unknown flaw that let the attacker partly escape the filesystem isolation around a build. From there the attacker obtained a build credential, pulled a copy of Railway’s internal monorepo, found a misconfigured OIDC binding in a cloud environment, and eventually reached a machine holding a credential that could decrypt customer environment variables. Railway’s database logs show queries that selected variables whose names matched attacker-supplied patterns for AI provider keys, and that favoured customers storing large numbers of variables. Railway says at most 1.85% of workspaces, about 100,000, could have had variables decrypted, and calls that an upper bound because its logs show which environments were exposed without showing individual variables. Decrypted data also included linked-account tokens for 1,852 users, covering bring-your-own agent keys and GitHub OAuth tokens. Railway reset all of them and found no logins using them. Its forensic team believes the AI keys were sold on to what it calls “distributed transfer stations” for use in model distillation.
Six AI Providers Gave Four Different Answers on Refunds
Railway’s report sets out each provider’s position as of November 10. AWS Bedrock and Google Gemini have started refunds through their account teams under existing fraud policies. OpenAI and SpaceXSI are reviewing claims case by case through a shared channel with Railway, and OpenAI asks affected customers to open a support case first. Anthropic is still in discussion with Railway, which is covering Anthropic-related losses in the meantime. OpenRouter has chosen to enforce the no-refund policy in its terms of service, and Railway says it will cover those costs too. Customers file a single charge report form with Railway to claim. The Hacker News, which reported the incident on November 11, noted that the form states Railway “is not responsible for usage of these keys” even as the report commits Railway to covering losses, and that neither document sets a cap or a deadline for claims.
Railway Changed Its Refund Position Within Two Weeks
On October 26, Railway told customers asking about refunds that it was unable to refund spend made with compromised keys and pointed them to the AI providers, The Hacker News reported. The November report reverses that and makes Railway the backstop. The Hacker News also points out that the report does not say how the $740,000 figure was estimated or what period it covers. Railway places the incident in a wider pattern. It names OpenCode, Warp and Fireworks AI among companies that disclosed similar thefts of AI credentials between October 6 and 23, and says the misuse it has seen ranges from coding-agent requests to AI training pipelines. That pattern fits credentials being relayed or resold to buyers far removed from the original intrusion. Railway has filed a complaint with the FBI’s Internet Crime Complaint Center and has not named or attributed the attacker.
Your Refund Rights for Stolen AI Keys Are Set Before the Breach
The useful finding in Railway’s list is that the customer’s outcome was fixed by contract terms signed long before October 20. Two providers treated the spend as fraud under policies that already existed. One provider pointed to terms of service that rule out refunds. The rest decided case by case. None of that depended on how carefully the customer had behaved, because the key was taken from the hosting platform’s database. Railway stepping in is a choice the company made, and its own claim form disclaims responsibility. A team whose key had sat on a platform less willing to pay would have absorbed the OpenRouter charges itself. Filing a claim also does nothing to stop the bleeding. The Hacker News notes that a charge report does not revoke a key, so the key has to be rotated with the provider that issued it.
Why This Matters for Engineering Leadership
Most engineering organisations manage AI keys as configuration. Railway’s report shows they behave like payment instruments, with unlimited spend attached and a fraud policy set by each issuer. Treat them that way. Start with an inventory: every AI provider key in use, which service owns it, and every platform that stores a copy, including hosting providers, CI systems and agent tools. Railway had to send providers hashes of every key it held because it could not yet tell which were exposed, and a team with its own inventory can rotate in minutes instead of waiting for that notice. Issue a separate key per service and environment, so one leak has a small blast radius and the unauthorised bill is easy to separate from legitimate usage. Set spend limits and usage alerts at each provider where they are offered, sized to what the service actually uses. Then read the fraud and refund terms of every provider and reseller you route through, and record them next to the inventory. In this incident, the same stolen key was worth a refund at one provider and nothing at another. If a gateway or reseller in your stack has a no-refund clause, decide now whether its convenience is worth carrying that exposure, and ask your hosting platforms in writing what they will cover if secrets they store are stolen.
FAQ
How much did the stolen AI keys from Railway cost customers?
Railway estimates unauthorised spend on the stolen keys reached about $740,000. Its November 10 incident report does not say how that figure was calculated or what period it covers.
Which AI providers are refunding Railway customers?
AWS Bedrock and Google Gemini have started refunds under their existing fraud policies. OpenAI and SpaceXSI are reviewing claims case by case, Anthropic is still in talks, and OpenRouter is enforcing its no-refund terms.
Will Railway pay for AI charges that providers refuse to refund?
Yes. Railway says it will cover token losses wherever a provider will not refund them, including OpenRouter costs and, for now, Anthropic costs. On October 26 it had told customers it could not refund that spend.
How many Railway workspaces were affected?
Railway says the attacker could have decrypted environment variables for at most 1.85% of workspaces, about 100,000. Its logs show which environments were exposed without listing individual variables, so that number is an upper bound.
Does filing a Railway charge report revoke a stolen key?
No. A charge report only starts a refund claim. The key still has to be revoked or rotated with the AI provider that issued it.
What should engineering teams do to limit losses from stolen AI keys?
Keep an inventory of every AI key and every platform that stores a copy, issue separate keys per service, set provider spend limits and alerts where offered, and read each provider’s fraud and refund terms before an incident.
