Railway says attackers used AI agents to breach its build systems and steal AI keys
Railway's incident report says attackers used AI agents to breach its systems and steal AI keys, running up about US$740,000 in fraudulent usage.
Cloud hosting company Railway has published its full report on the October security incident it first disclosed last month. The report, dated 10 November, says the attackers used AI agents to find exploits and move through Railway's systems quickly. It also says the stolen AI provider keys were used for about US$740,000 in unauthorised usage.
What Railway says happened
According to Railway's report, the attacker started creating dummy accounts on 18 October and began probing the build infrastructure on 20 October. The main steps were:
- Build system: a previously unknown flaw in Railway's shared build infrastructure let the attacker partly escape filesystem isolation and obtain a build credential. Railway says an internal system accepted that credential, exposing the histories of 353 builds.
- Source code: build cache snapshots gave the attacker a copy of Railway's internal monorepo.
- Cloud and secrets: on 21 October, a misconfigured OIDC binding gave the attacker a short-lived cloud credential. From there the attacker reached a longer-lived decryption credential, then queried and decrypted customer environment variables, prioritising AI provider keys.
Railway says the attacker's access had ended by mid-morning UTC on 21 October. It then invalidated all user sessions and tokens and restored builds in stages over the following day.
The role of agents
Railway says the attacker ran fanouts of agents over its monorepo to find exploits and move laterally, and that the evidence suggests they used abliterated models, meaning models with their safeguards removed. Railway described its response as a contest of agent speed. Citing the security firm Wiz, it said it was defending against those models and a harness similar to ARTEX, the penetration-testing framework named in the earlier WorkOS incident.
Railway also says that after the incident it obtained access to Project Glasswing, with approval from the US Treasury.
Scale and cost
Railway says up to 1.85% of workspaces, about 100,000, may have had environment variables decrypted. It stresses that this is an upper bound based on environment-level logs, not a confirmed count. It reset linked-account tokens for 1,852 users and says it found no logins using stolen credentials.
Railway believes the AI keys were sold on to resale services for use in model distillation. It has not established whether the original attacker ran those services. It says it is seeking refunds from providers and is covering some costs itself, and that it has filed a complaint with the FBI's IC3.
A wider pattern
BleepingComputer reports that Wiz links the incident to a wider campaign it calls Black October, which has also hit Northflank, OpenCode, Warp and Lightning AI. Railway's own report says OpenCode, Warp and Fireworks AI reported similar thefts between 6 and 23 October. The two sources list slightly different companies.
This breach is separate from the episode tracked earlier this month, in which a Railway engineer's testing agents probed rival host Primcloud. Railway's report does not mention that episode.
What Railway has changed
Railway lists these fixes:
- replacing shared build machines with a new short-lived virtual machine for each build
- rotating credentials and rebuilding its certificate infrastructure
- adding two-factor and biometric approval for critical CI actions
- moving towards short-lived credentials and centralised audit logging of encryption operations
It has also set up a form for customers to report unauthorised charges.
The incident is a case of human attackers using agents as tools, not an agent acting beyond its operator's instructions. It suggests that agent-assisted attacks can compress the time defenders have to respond. Railway's on-call engineer was paged about the first anomalous activity roughly six and a half hours after the initial exploit.
Sources
- Railway: Incident Report: October 21st Security Attack
- BleepingComputer: Railway breach leads to $740,000 in fraudulent AI charges