Recorded from the disclosing organization's own preliminary report. Figures below are the figures published in that report. Railway states that a final forensic report will follow, and this entry is marked for revision until it does.
A build path reached the keys that unlocked everything else.
Railway detected abnormal activity at 13:58 UTC on 20 October and declared an incident at 17:03, halting build pipelines. Over the following day it found further escalation paths and additional attacker access, restricted credentials and halted builds a second time. It confirmed the attacker no longer had access at 03:32 UTC on 22 October.
Railway's account describes an attempted privilege escalation in build infrastructure, from which the attacker reached CI credentials, decrypted artifacts holding further API access, queried the database for encrypted environment variables, and decrypted some of them using the compromised credentials. Railway states the attacker's objective was obtaining AI provider API keys.
About 1.85% of workspaces — roughly 100,000 — may have had environment variables compromised. Railway says the number of affected users is larger, because reset requests went out for all stored keys.
Keys proactively revoked spanned source control, model providers and payment services. Storage credentials, application installations and platform access tokens were rolled or revoked.
Railway states user accounts were not logged into, and that it has not ruled out access to user source code. The initial entry point is not stated in the preliminary report.
Railway attributes the sophistication of the exploits to the attacker's access to unrestricted models and substantial compute. That is the company's own characterization of capability, not a documented finding about the attack path, and it is recorded here as such.
The secrets were encrypted. The key was in the same building. Environment variables were held encrypted at rest, and the credentials capable of decrypting them were reachable from the build infrastructure the attacker had already compromised. Encryption that shares a blast radius with the credential that unlocks it defends against a stolen disk and against nothing that happens after a foothold.
Behind that, credentials were long-lived and scoped past their purpose — a build credential that worked on internal systems it had no reason to reach — and a single escalation in shared build infrastructure did not stop at one customer's boundary. Railway states its design is intended to ensure one compromised key set does not compromise the whole system, and its report does not explain how that intent was defeated.
The finding that transfers has nothing to do with Railway. Every organization using the platform inherited an incident it had no part in, and could only respond at the speed of its own inventory. The ones who rotated quickly were the ones who already knew what they had stored, where, and which of it mattered. Rotation is not the hard part. Knowing the list is the hard part, and it has to exist before the notification arrives.
Exposure floors are recorded, not averaged down.
An organization that has placed model provider keys, database URLs or payment credentials into a hosting or build platform carries this structure directly. The analogue is heightened because the credentials concerned are the ones that unlock other systems, and because the organization holding the exposure cannot see the platform's internal compromise, cannot influence its containment, and learns the scope on the platform's schedule.
Risk & control readiness
Vendor and platform review, secret inventory and ownership, credential scope and lifetime, rotation that can be executed rather than described, and incident readiness for compromises that originate outside the organization.
Could we list today, without asking anyone, every credential we have placed in a third-party platform?
Then ask how long rotating them would take. A list that takes a week to assemble is the same as no list, because the notification does not wait.