Railway Cyberattack 2026: Hackers Steal AI API Keys, Trigger $740,000 in Fraudulent Usage

Railway, the cloud platform developers use to deploy applications, says an attacker who escaped its build isolation on October 20, 2026 went on to decrypt customer environment variables for at most 1.85% of workspaces, about 100,000, while hunting for AI provider API keys. Those keys have since generated an estimated $740,000 in unauthorized usage. The figures come from Railway’s detailed incident report of November 10, 2026; the 100,000 figure is an upper bound, not a confirmed count of stolen secrets.

Incident overview: what happened at Railway

Railway became aware of unauthorized access at 13:58 UTC on October 20, 2026, according to its incident report. The attacker chained several exploits, moved laterally across internal systems and decrypted a subset of customer environment variables. Railway says the attacker was fully evicted at 09:24 UTC on October 21 and had no access to customer data after 09:25 UTC.

The attack was global in scope because Railway is a multi-tenant platform-as-a-service. Customers deploy applications on it and store secrets such as API keys, database credentials and registry credentials as environment variables. Railway’s forensic work indicates AI provider keys were the primary target.

Key confirmed figures from Railway’s report:

  • Up to 1.85% of workspaces (about 100,000) could have had environment variables decrypted. Railway says its logs identify environments, not individual variables, so this is an upper bound.
  • 1,852 users had linked-account tokens (BYOK agent keys and GitHub OAuth tokens) decrypted. All were reset.
  • 353 build histories were read, including variables and secrets passed to those builds.
  • About $740,000 in unauthorized AI usage on stolen keys is estimated. Railway has not published how the estimate was calculated or what period it covers, as The Hacker News noted.
  • Internal source code was exposed. The attacker obtained a copy of Railway’s internal monorepo. Railway says it has no indication that customer repositories were downloaded or customer databases accessed.

Running customer services stayed online throughout. The disruption fell on builds, deployments, logins and API access, which were paused or restricted in stages.

Timeline of the Railway breach

The timeline below is condensed from Railway’s incident report, newest first. Railway notes it is not exhaustive and that gaps do not mean inactivity.

Date and time (UTC)Event
Nov 10, 2026Railway publishes its detailed incident report, revising times from the October 22 preliminary report.
Nov 4, 17:30SSH access restored for services using static egress.
Nov 2New signups reopen after being paused since Oct 22, 19:06.
Oct 22, 22:49Builds restored for all plans, including Free and Trial. Enterprise builds returned at 03:32, Pro at 06:29, Hobby at 21:08.
Oct 22, 01:22Railway publishes credential-rotation and usage-check guides before builds return.
Oct 21, 20:19All user sessions and tokens invalidated, including CLI, mobile and third-party apps.
Oct 21, 09:25Railway begins notifying GitHub, GitLab, Anthropic, OpenAI, OpenRouter, SpaceXSI, Google and Stripe with hashes of every key it held.
Oct 21, 09:24Attacker fully evicted.
Oct 21, 07:14–07:22Cloud credentials disabled, static credentials revoked, key decryption service shut off, builds halted again.
Oct 21, 06:36–07:15Second window of variable decryption through the key management service.
Oct 21, 04:42–05:14First window of decryption. Encrypted variables were pulled from the database between 04:07 and 05:21.
Oct 21, 03:26Attacker obtains a short-lived cloud credential through a misconfigured OIDC binding.
Oct 20, 17:28Compromised build credential revoked and replaced; builds had been halted at 17:03.
Oct 20, 13:58On-call engineer paged by anomalous calls to the build-artifact registry. The attacker had read 353 build histories between 13:50 and 15:30.
Oct 20, 07:28Attacker exploits a previously unknown build-isolation flaw to partially escape filesystem isolation. Probing began at 06:50.
Oct 18Attacker begins creating dummy accounts; early attempts failed against automated defenses.

Root cause: four weaknesses that compounded

Railway’s report does not blame a single flaw. A postmortem analysis from The Cyber Security Guru summarizes the pattern well: a build isolation bug, an over-trusted internal credential, a misconfigured OIDC binding and a decryption credential on a reachable disk were each containable alone. Together they formed a path from a customer build to the secrets database.

1. Build isolation failure

Railway’s build system ran multiple customer builds on the same virtual machine, each in supposed full isolation. A previously unknown vulnerability let the attacker partially escape filesystem isolation and read some files on the host. Railway has not published a CVE identifier or a technical description of the flaw, and none is known to this article.

2. An over-trusted build credential

The stolen credential was meant to authorize build work such as pushing images to a registry. Railway’s self-hosted Temporal cluster, which drives parts of the build process, nonetheless accepted it as a trusted client. That allowed the attacker to call Temporal client APIs and read stored build histories, which contained environment variables and secrets passed to builds.

3. Cloud identity misconfiguration

After scanning the stolen code, the attacker found an erroneously defined OpenID Connect (OIDC) binding in one of Railway’s cloud environments. The binding is part of how the CI/CD system obtains short-lived cloud credentials. Exploiting it at 03:26 UTC on October 21 gave the attacker a short-lived credential to production cloud infrastructure.

4. A long-lived decryption credential on disk

An orchestration machine that provisions customer workloads stored a longer-lived credential on disk that could authenticate to the key management service. The machine also sat on the internal network that reaches the database of envelope-encrypted variables. Railway has since restricted decryption so that only its secrets service can reach keys that unwrap data encryption keys.

Contributing factor: source-code exposure

The build credential could fetch repository snapshots used for caching. That let the attacker retrieve a copy of Railway’s internal monorepo. Railway says this gave the attacker a privileged view of the platform and enabled fast, parallel exploration. The company only realized the attacker had the code when it saw probing of endpoints that had not been live for weeks, months or years.

Attack methodology: from build sandbox to decrypted secrets

The sequence below reflects Railway’s own account. Railway describes the escalating situation as a contest of agent speed, and says it defended against what it believes were unrestricted (“abliterated”) models using a harness similar to ARTEX, according to Wiz. That is Railway’s assessment, not an independently verified finding.

  1. Reconnaissance and account staging (from Oct 18). The attacker created swarms of dummy accounts. Railway’s automated systems blocked early attempts.
  2. Build isolation escape (Oct 20, 07:28 UTC). A zero-day in the multi-tenant build infrastructure exposed host files, including a credential used to authenticate to some internal systems.
  3. Internal API abuse. The credential was accepted by Temporal. The attacker read stored histories of other builds, including 353 read between 13:50 and 15:30 UTC.
  4. Source-code theft. A recent build of Railway’s support platform led to repository snapshots, from which the attacker retrieved the internal monorepo.
  5. Re-entry from new angles. After Railway revoked the build credential and re-enabled builds, the attacker tried to escape isolation again and failed, then used the code to probe other routes.
  6. Cloud identity abuse (Oct 21, 03:26 UTC). A misconfigured OIDC binding returned a short-lived cloud credential. The attacker used it to fetch two internal credentials: one able to read encrypted variables from a production database inside the internal network, and an SSH private key used by the CD system.
  7. Lateral movement. The SSH key opened an orchestration machine. A file on that machine held a longer-lived credential permitting variable decryption.
  8. Targeted collection (04:07–05:21 UTC). Database logs show queries selecting rows whose variable names matched attacker-supplied regular expressions for AI provider keys. The attacker also prioritized customers with many variables.
  9. Decryption (04:42–07:15 UTC, two windows). The attacker used the orchestration machine’s credential against the key management service to decrypt a subset of the retrieved variables.

Railway’s variable storage uses envelope encryption: each variable is encrypted with a per-environment data encryption key, which is itself encrypted by a root key held in an upstream cloud provider’s key management service. The root key cannot leave the service, which limited the attacker to decrypting what the compromised credential allowed.

An additional database query also matched cloud, database and payment-service credentials. No ransomware or malware family has been identified in Railway’s public account.

Threat actors and attribution: what is confirmed and what is not

No threat actor has been publicly named. Railway has not identified the attacker, has filed a complaint with the FBI’s IC3 and is working with a specialist forensics firm whose investigation is ongoing. It is sharing indicators of compromise privately for now and plans to publish them once law enforcement signs off.

Railway’s assessment (stated belief, not proven attribution):

  • The attacker was specifically targeting AI provider credentials, based on the variable-name queries in database logs.
  • The credentials were likely passed to services that relay or resell AI provider access, for use in model distillation.
  • Observed unauthorized usage ranged from coding agents to AI training pipelines, according to customer-shared logs and IP addresses. Railway says this is consistent with credential-resale services but it has not established whether the original attacker operated them or supplied other parties.

Wider context from Wiz. Wiz’s Black October analysis counts at least nine AI developer platforms that disclosed incidents between October 6 and October 30, 2026, including Fireworks AI, Zeabur, Lightning AI, OpenCode, Railway, Warp and Northflank. Wiz says all public reports describe attackers hunting stored environment variables, especially those granting AI access, but that no single vulnerability or dependency links the attacks. Railway says Wiz’s research provides context and does not independently attribute the Railway incident or show that all later misuse came from one operator.

On the “ARTEX” reference. Wiz documented an ARTEX page described as an autonomous penetration-testing console, and noted that related infrastructure was publicly named by Zeabur in its own compromise. Railway says it defended against a harness similar to ARTEX, citing Wiz. That is a tooling similarity, not evidence of who operated the Railway intrusion.

Until law enforcement or the forensics firm publishes findings, readers should treat any claim tying Railway’s breach to a specific country, group or AI lab as unverified.

Impact assessment: financial, operational and regulatory

Financial impact and who pays

Railway estimates unauthorized spend on stolen keys at about $740,000 and says it will refund via providers where possible and cover costs from its own balances where providers refuse. Provider positions as of the November 10 report:

ProviderPosition per Railway’s report
AWS BedrockExisting fraud refund policy applies; account teams have initiated refunds.
Google GeminiExisting fraud refund policy applies; account teams have initiated refunds.
OpenAICase-by-case review; customers open a support request and pass the case ID to Railway support.
SpaceXSICase-by-case review via a shared bridge with Railway.
AnthropicIn active discussions; Railway is covering Anthropic breached-key costs in the interim.
OpenRouterEnforcing its no-refund terms; Railway says it will cover all OpenRouter-related costs.

The report does not say how the $740,000 was estimated, over what period, or whether any cap or claim deadline applies, according to The Hacker News. Customers report charges through a form Railway published. One customer-reported example cited by Sesame Disk from Railway’s community forum described a Gemini key flagged by Google at EUR 133.12 of usage against a predicted EUR 0.02 baseline; it is an anecdote, not a Railway figure.

Operational impact

Running services were not affected by the build pause, but customers could not deploy for roughly two days. Revoking a credential could break running services that depended on it, and because builds were unavailable some customers could not redeploy with a replacement credential. Railway acknowledges that some customers experienced downtime for this reason. Other effects included invalidated API and project tokens, invalidated sessions, storage-bucket credential rotation that paused database backup archiving for some customers, and a signup freeze from October 22 to November 2.

Data exposure

Exposure was concentrated in environment variables, linked-account tokens and Railway’s own source code. Railway reports no evidence of customer repository downloads or customer database access, and no logins to Railway accounts using stolen credentials. Any data copied but not decrypted cannot now be decrypted, because Railway disabled the credential the attacker used.

Reputational and regulatory dimensions

Railway chose to notify AI providers before customers, arguing that a broad notification would delay eviction and trigger traffic surges that slowed recovery (it reports nearly two orders of magnitude above baseline build capacity). Railway concedes it could have been faster and says it will set a fixed period for assessing impact after which broad notification occurs regardless. Railway says it is sharing IOCs with customers that have compliance reporting duties. No regulator action or fine has been reported as of this writing.

IT, OT, IoT and industrial cybersecurity implications

Nothing in Railway’s report indicates that industrial control systems, SCADA networks or physical processes were touched. The relevance to OT is indirect but real, and the following is OT Ecosystem’s analysis, not a finding reported by Railway.

Industrial software increasingly runs on platforms like this. Many asset-monitoring dashboards, telemetry ingestion services, historian front ends, maintenance portals and vendor integrations are built as cloud-hosted web applications. Their deployment platforms hold the secrets that connect them to data sources: database credentials, message-broker keys, cloud storage keys, webhook secrets and increasingly AI service keys. A platform compromise exposes those secrets in bulk, regardless of how well the application itself is written.

Environment variables are a single point of failure. Railway’s attacker used simple regular expressions over variable names and targeted customers with many variables. Any organization that keeps credentials for historians, remote-access gateways, IoT device fleets or industrial data lakes in a hosting platform’s variable store faces the same pattern. Railway’s own report notes that one query also matched cloud, database and payment-service credentials, though it found no evidence that customer databases were accessed.

The control-plane lesson mirrors OT segmentation. The attacker moved from a customer-controlled build to an internal workflow engine to a cloud identity to a decryption service. Each hop crossed a boundary that was assumed to be isolating. OT defenders will recognize the pattern: shared services, over-broad service accounts and stale trust relationships are what let an intruder cross zones and conduits.

Agent-speed attacks shrink detection windows. Railway reports the attacker probed many routes in parallel and moved quickly after obtaining source code. Wiz separately reports that agentic tooling helps attackers compromise victims faster. Industrial teams that rely on slow patch cycles and manual review should plan for shorter exploitation timelines against internet-reachable engineering and monitoring portals.

IoT and AI-enabled telemetry. Teams that embed AI services into condition monitoring, anomaly detection or operator-assistance tools should treat the provider key as a financial and data-access credential. A stolen key can drive large bills and, depending on how the application is built, expose prompts and data sent to the provider. Railway’s report does not say whether prompts or responses were accessed.

Incident response and recovery

Railway’s on-call engineer was paged by an anomaly alert on a registry service about six and a half hours after the first exploit, then traced the calls to a compromised build credential and the escaped build. An incident response team was formed, a fix was merged and builds were halted until it was rolled out. Railway says it observed the attacker around the clock from that point.

Containment steps reported by Railway:

  • Halted builds twice (Oct 20, 17:03 and Oct 21, 07:22 UTC) and revoked and replaced the build credential.
  • Disabled all active cloud credentials, revoked static credentials and shut off the key decryption service at 07:14 UTC on October 21.
  • Invalidated every user session and token, rotated storage bucket credentials and reset GitHub tokens it held for customer accounts.
  • Notified AI and developer-platform providers with hashes of every key it held so they could revoke quickly. Some providers revoked further keys at their discretion, which also revoked some replacement keys customers had already issued.

Eradication and rebuild. Railway bootstrapped new PKIs and a new root CA, rotated every relevant key and secret, rebuilt internal tooling from validated source and reprovisioned suspect machines, snapshotting the old ones for forensics. CI was disabled during rotation, and each dependent service was restarted manually.

Recovery status as of November 11. Builds have been restored for all plans since October 22 at 22:49 UTC, signups reopened on November 2 and support tooling returned on October 26. Railway says every attack vector used is closed and that it has gained access to Project Glasswing with approval from the U.S. Department of the Treasury. The forensic investigation, the law enforcement inquiry, provider refund discussions and public release of indicators of compromise remain open.

Hardening Railway reports:

  • Each build now runs in a dedicated ephemeral virtual machine with a credential limited to pushing one specific image to the registry.
  • CI infrastructure was rebuilt with fine-grained access control, 2FA and biometric approval in critical areas, network isolation and SIEM alerting.
  • Service accounts were reviewed for least privilege, short-lived just-in-time credentials were adopted where possible and deprecated systems were decommissioned.
  • All encrypt and decrypt operations are now centrally audit-logged, no service other than the secrets service can reach keys that unwrap data encryption keys, and each caller can decrypt only the credentials it needs.

Security recommendations

For Railway customers and anyone who stored secrets on a PaaS

  1. Rotate at the issuer, not just on the platform. Deleting a variable in a hosting dashboard does not invalidate the key. Revoke and reissue AI, cloud, payment, database and registry credentials with the provider that issued them. Railway advises rotating anything stored before October 21, and a provider revocation notice alone does not prove your key was stolen.
  2. Review provider usage and billing. Check usage logs for your AI, cloud and payment providers from October 18 onward. Preserve logs before deleting suspicious resources. Report unauthorized charges through Railway’s form and file claims with the provider.
  3. Set spend limits and per-key scope. Use separate keys per application and environment, set hard budgets and quotas, and alert on usage anomalies. A key that can only call one model within a budget limits the damage from theft.
  4. Reduce exposure of databases. Remove public network access to databases, and review GitHub audit logs, authorized apps and registry audit logs (for example Docker Hub), as Railway recommends.
  5. Keep secrets out of long-lived static variables. Prefer short-lived, workload-bound credentials and a dedicated secrets manager with audit logging for high-value keys.
  6. Plan for redeploy during an incident. Railway customers who revoked build credentials could not redeploy while builds were paused. Maintain a tested path to rotate credentials without a full rebuild.

For platform, cloud and DevSecOps teams

  1. Treat multi-tenant builds as hostile execution. Run each build in its own ephemeral microVM or VM, with no shared host files and no credentials beyond what that one build needs.
  2. Scope workflow-engine trust. A credential intended for registry pushes should not be accepted by an orchestration API such as a Temporal cluster. Enforce per-client authorization, namespaces and network paths.
  3. Audit OIDC trust policies. Restrict cloud role trust to exact issuer, audience and subject claims (repository, branch, environment), and test negative cases. A permissive binding turns a CI identity into a cloud entry point.
  4. Remove decryption material from reachable disks. Centralize encrypt and decrypt operations behind one audited service, restrict each caller to the keys it needs, and alert on bulk decrypts or unusual key management service use.
  5. Minimize what a stolen build can reach. Do not place secrets or full repository snapshots in task histories, and decommission dead endpoints and unused services, since attackers use leaked source to find them.
  6. Detect at machine speed. Alert on anomalous registry and internal API calls, mass variable-name queries and unfamiliar account swarms, and rehearse isolating an attacker while observing them.

For incident response and communications

  1. Decide notification order in advance. Railway notified providers before customers and later acknowledged the cost in trust. Define when providers, affected customers and everyone else are told, and set a fixed deadline for broad notification.
  2. Anticipate rotation surges. Mass credential rotation produced build traffic nearly two orders of magnitude above baseline for Railway. Plan capacity and staged guidance.
  3. Expect over-revocation. Providers may revoke replacement keys issued during the incident, so state clearly which new keys are safe.

Leave a Reply

Your email address will not be published. Required fields are marked *