On October 20 and 21, an attacker escaped Railway’s build isolation, moved laterally through its cloud, and decrypted customer environment variables for at most 1.85% of workspaces, about 100,000 (Railway). Railway says the attacker used abliterated models with no safeguards and, according to Wiz, a harness similar to ARTEX. Its queries targeted AI provider keys, and unauthorized spend on stolen keys reached an estimated $740,000.
What happened
- Oct 18The attacker begins creating a swarm of Railway accounts
- Oct 20, 07:28A build partially escapes filesystem isolation and reads a build server credential from the host
- Oct 20, 13:50The credential reads 353 build histories from Railway’s Temporal cluster and fetches a copy of the monorepo
- Oct 20, 13:58Anomalous registry API calls page the on-call engineer
- Oct 21, 03:26A misconfigured OIDC binding hands the attacker a short-lived cloud credential
- Oct 21, 04:07From an orchestration machine, the attacker pulls encrypted variables, prioritizing AI key names
- Oct 21, 04:42Decryption through the key management service begins and runs until 07:14
- Oct 21, 09:24Railway evicts the attacker
Railway believes the stolen keys were sold to services that resell model access for distillation. It notes that OpenCode, Warp, Fireworks AI, and others disclosed similar thefts between October 6 and 23. Fireworks links its own incident to a financially motivated LLMjacking group.
The monorepo copy let the attacker probe endpoints with no live references and fan out agents across escalation paths. One path was a CI/CD identity binding that issued cloud credentials, which led to the credentials and SSH key needed to reach the variables database and decrypt its envelopes. Railway describes the response as a contest of agent speed.
How it could have been caught
Build machines push images and talk to a short list of services, and an orchestration host follows a fixed provisioning routine. Each step in this intrusion made one of those machines do something new. A baseline would have flagged the first host read from inside a build at 07:28, six and a half hours before the page.
What this looks like in Trace
Here is the intrusion traced across Railway’s infrastructure. The dates, times, systems, and credential types come from Railway. The process ancestry, hostnames, and ports are illustrative.
Network trace. Workloads build/vm-pool and fleet/orchestrator. Each row is a tuple Trace’s sensor recorded, checked against the baseline.
- Oct 20, 07:28buildkit → sh → cat (host filesystem)
- Baseline
- Anomaly, build reads host files
- Oct 20, 07:40build worker
- Destination
- registry:443
- Baseline
- Known
- Oct 20, 13:50sh → curl (build credential)
- Destination
- temporal-frontend:7233
- Baseline
- Anomaly, client API from a build host
- Oct 21, 04:07sshd → bash → psql
- Destination
- variables-db:5432
- Baseline
- Anomaly, interactive session
- Oct 21, 04:42bash → python3
- Destination
- cloud key management:443
- Baseline
- Anomaly, same case
The attacker’s models ran on its own infrastructure, so no AI proxy would have seen them. The stolen keys were used from outside Railway, where the theft itself is the evidence.
Traceback. At October 21, 04:07, the triage agent adds to the open case and follows the chain:
decrypt calls → psql over SSH on the orchestration host → SSH key from the cloud environment → short-lived credential from an OIDC exchange with no CI run → misconfigured binding
The chain ends at a credential no pipeline requested, and that gap is the finding.
Verdict: an outside agent chaining a build escape into decryption of customer secrets. The case reaches Railway’s security team at 07:28 on October 20, with the build, the host read, and the credential named, nearly a day before the first decryption.
This analysis reconstructs the incident from public reporting and walks through where Trace would have raised the first alert inside Railway’s infrastructure. Detection times are illustrative.