What the attacker was actually after

The hosting platform Railway published its full report into an October security incident on 10 November. The figure that stands out is the bill: unauthorised spend on keys taken during the attack was estimated to reach roughly $740,000.

Railway says its forensic investigation concluded the attacker was specifically targeting AI provider credentials, to sell to what the report calls distributed transfer stations for use in distillation. That is a different motive from the usual ransomware or data-theft pattern, and it puts a price on a company’s API keys that has nothing to do with the data behind them.

Coiled fibre optic cables on a rack
The attacker reached customer environment variables through an internal network. Illustrative photo. Brett Sayles · pexels · Pexels License

The attacker reached and decrypted a subset of customer environment variables for at most 1.85% of workspaces, which Railway puts at around 100,000. Among the decrypted data were linked-account tokens — agent keys and GitHub OAuth tokens — for 1,852 users. Railway says it reset every one of those tokens, notified each user, and found no logins to Railway accounts using credentials taken in the incident.

A contest measured in hours

The timeline in the report is unusually tight. The attacker began probing Railway’s build infrastructure at 06:50 UTC on 20 October and exploited a previously unknown vulnerability in it at 07:28, partially escaping filesystem isolation and obtaining a build credential. An alert on anomalous API calls paged the on-call engineer at 13:58. The attacker was fully evicted at 09:24 UTC on 21 October.

Railway describes it as “not a single breach followed by a single fix” but a contest of agent speed, and says the attacker was using abliterated models — models with their safeguards stripped — able to perform escalated actions in rapid succession. It attributes to the security firm Wiz the observation that the attacker’s harness resembled ARTEX.

The second entry point is the one worth reading closely. Having obtained a copy of Railway’s internal monorepo through stored build histories, the attacker scanned the code and found a misconfigured OIDC binding in one of Railway’s cloud environments, which yielded short-lived deployment credentials and, from there, a credential that could decrypt customer variables.

A close-up of a circuit board
Unauthorised spend on the stolen keys was estimated at about $740,000. Illustrative photo. Miguel Á. Padriñán · pexels · Pexels License

Who pays

The refund section is the part that will matter to anyone running keys on someone else’s platform. Railway says AWS Bedrock and Google Gemini have existing fraud refund policies and their account teams have initiated refunds. OpenAI and SpaceXSI are reviewing case by case through a shared bridge with Railway support. Anthropic was still in discussion at the time of writing, and Railway says it will cover those costs itself in the interim. OpenRouter is enforcing its no-refund terms, and Railway says it will cover all OpenRouter-related costs.

So a customer’s exposure depends less on their own security than on which model provider they happened to be using, and on their host’s willingness to absorb the difference. Railway says every attack vector used has been remediated, that each build now gets a dedicated ephemeral virtual machine, and that it is working with upstream vendors on responsible disclosure for third-party vulnerabilities it found. The next thing to watch is whether the providers that have not yet set a policy here write one.