Identity · AI Security
Hackers stole AI credentials to rack up $740,000 in cloud charges
Attackers used stolen AI credentials from Railway to generate $740,000 in fraudulent compute charges. The company is reimbursing affected customers and has reset compromised keys.
Ross Grant3 min read
The short version
Hackers broke into Railway, a cloud hosting service, and stole credentials used to access AI services. They used these stolen keys to run expensive AI tasks, leaving Railway and its customers with a $740,000 bill. The company has reset affected keys and is paying back customers. If you used Railway, you should change your credentials and check for unusual charges.
By the numbers
- $740,000fraudulent AI compute charges generated by attackersCSO Online
- 100,000workspace credentials potentially at riskCSO Online
- 1,850+linked account tokens exposed and resetCSO Online
- 26 hoursduration of the attack before evictionCSO Online
What happened?
Hackers stole AI provider credentials from Railway, a cloud hosting company, and used them to generate $740,000 in fraudulent compute charges. The attack lasted nearly 26 hours before Railway evicted the intruders. The company is reimbursing affected customers and has reset compromised keys. Only CSO Online has reported the incident so far.
Who is affected
Railway, a cloud hosting provider, reported that attackers accessed and decrypted a subset of customer environment variables. According to CSO Online (opens in a new tab), up to 100,000 workspace credentials were at risk, which is about 1.85% of all workspaces. Over 1,850 users had their linked account tokens exposed, including GitHub and OAuth tokens. The company worked with several partners, such as GitHub, GitLab, Anthropic, OpenAI, and Stripe, to revoke potentially compromised keys.
How the attack worked
The attack began on 20 October 2026, when hackers exploited an unknown flaw in Railway’s build infrastructure. They bypassed filesystem isolation, obtained an internal credential, and accessed stored build histories. From there, the attackers retrieved Railway’s internal source code and analysed it for further weaknesses. They found another vulnerability, this time involving a misconfigured trust relationship, which let them decrypt customer environment variables. The attackers then used these credentials to generate fraudulent AI usage, chaining exploits and moving laterally across systems for nearly 26 hours before being evicted on 21 October. During this time, they racked up $740,000 in unauthorised compute charges.
What was taken
The attackers obtained and decrypted a subset of customer environment variables, including AI provider keys, cloud, database, and payment service credentials. They also accessed tokens for linked accounts, such as GitHub and OAuth. Customers reported that their AI keys were used to answer coding questions and perform training-data tasks they did not initiate. The breach exposed how stolen machine identities can be used to fuel expensive AI operations, such as model distillation (where knowledge from a large AI model is transferred to a smaller one).
The response
Railway disclosed the incident on 22 October and published a follow-up report. The company advised all users to revoke and reapply sensitive environment variables stored before 21 October. It reset compromised tokens and worked with partners to revoke affected keys. Internally, Railway shut down compromised tooling, rotated all keys, moved to a new generation of infrastructure, and added features to mass-revoke app permissions. The company is reimbursing customers for unauthorised charges, even where providers decline refunds. Customers are urged to check audit logs, rotate credentials, and set spending caps and usage alerts on AI keys.
What you should do
Anyone who stored secrets on Railway before 21 October should treat those credentials as compromised. Rotate all sensitive keys, especially AI provider keys, cloud, database, and payment credentials. Check usage logs for unexplained activity and alert Railway to any unauthorised charges. Limit the scope and spending of each AI key, and ask your provider how secrets are stored and who can decrypt them. Issue as few keys as possible and use short-lived credentials where available.
Why it matters
This attack shows how valuable AI credentials have become to criminals. Stolen machine identities can be used to run costly AI operations, with the real owner left to pay the bill. It highlights the need for stronger controls and monitoring of API keys and other credentials.
Key points
- Hackers stole AI credentials from Railway and used them for $740,000 in compute charges.
- Up to 100,000 workspace credentials and 1,850 account tokens were at risk.
- Attackers exploited two vulnerabilities to access and decrypt customer secrets.
- Railway has reimbursed customers and reset compromised keys.
- AI credentials are now a major target for criminals, like credit card numbers.
Sources
CSO Online
Reference


