Railway says stolen AI keys drove about $740,000 in unauthorized charges
Short answer
Railway’s November 10, 2026 incident report says an attacker who broke out of the build system on October 20 decrypted customer environment variables for at most 1.85% of workspaces, about 100,000, an upper bound rather than a count of confirmed stolen keys. The queries targeted AI provider credentials. Railway estimates unauthorized spend on the stolen keys at about $740,000 and says it will cover charges that providers will not refund. If a key lived in Railway before October 21, rotate it at the provider and review that provider’s usage.
Last updated · Sekrd Research
- When: probing from October 18, 2026; first exploit at 07:28 UTC on October 20; customer-data access ended by 09:25 UTC on October 21. Report published November 10, 2026.
- Who: Railway, a platform that builds and runs customer apps and stores their environment variables. Authors of the report: Jacob Cooper, Angelo Saraceno, Sarah Bedell, and Luca Casonato.
- Impact: variables in at most 1.85% of workspaces (about 100,000) could have been decrypted. Linked-account tokens for 1,852 users were reset. Unauthorized spend estimated at about $740,000. Histories of 353 builds were read.
- Root cause: a previously unknown isolation bug in multi-tenant builds, then an erroneous OIDC binding that yielded a short-lived cloud credential.
- Status: Railway says the attacker was fully evicted at 09:24 UTC on October 21, each exploited weakness is closed, and a forensics firm’s investigation is ongoing. No attacker name is in the report.
What happened
On October 21, 2026, Railway lost control of a slice of the secrets customers had stored there. The November 10 incident report is the document to read. It says engineers were paged at 13:58 UTC on October 20 after anomalous calls to a registry that stores build artifacts. By then an attacker already had a build credential, a copy of an internal monorepo, and a path toward the database that holds envelope-encrypted environment variables.
Logging shows that at most 1.85% of workspaces, about 100,000, could have had any environment variables decrypted. The logs identify environments, not individual variables, so the figure is an upper bound. Railway emailed customers it judged high risk. Later copies cannot be decrypted, because that credential was disabled. The queries targeted AI provider keys, and another matched cloud, database, and payment credentials. Forensics did not show customer databases being accessed, and Railway says customer repositories were not downloaded from the build system.
Timeline
- October 18. First attacker signups. Railway says the early attempts failed.
- October 20, 07:28 UTC. Partial filesystem escape on a shared build machine. A build credential was read.
- October 20, 13:50–15:30 UTC. Histories of 353 builds read, sealed variables included. On-call paged at 13:58 UTC.
- October 20, 17:03 UTC. Builds halted. The build credential was revoked at 17:28 UTC.
- October 21, 03:26 UTC. Short-lived cloud credential via an erroneous OIDC binding.
- October 21, 04:07–07:15 UTC. Encrypted envelopes pulled, then some of them decrypted. Queries preferred AI provider key names.
- October 21, 07:14 UTC. Cloud credentials and key decryption disabled. Builds halted again at 07:22 UTC.
- October 21, 09:24 UTC. Attacker fully evicted, Railway says. No customer-data access beyond 09:25 UTC. Providers notified from 09:25 UTC with hashes of every key Railway held.
- October 22. Guides at 01:22 UTC. Builds restored by plan from 03:32 UTC (Enterprise) through 22:49 UTC (Free and Trial).
Why the keys came out
Two mistakes stacked. A partial filesystem escape on a shared build machine produced a credential meant only for build work, such as registry pushes. An internal Temporal cluster treated it as a trusted client, so the attacker could read other builds’ histories and then an internal monorepo. Railway says that, up to that point, it has no indication the attacker was after anything but Railway itself. With the codebase, the attacker used an incorrectly defined OIDC binding at 03:26 UTC on October 21 to obtain a short-lived cloud credential, then an SSH key and a credential that could read encrypted variables from inside the network. An orchestration machine held a longer-lived decryption credential on disk.
Each environment encrypts its variables with its own data key, and that key is wrapped by a root key Railway says cannot leave the key management service. The attacker still needed a credential that could ask the KMS to unwrap keys. That is why the ceiling is a set of workspaces, not every customer, and why the ceiling is not a list of keys that were proven stolen.
Railway believes the credentials were passed on for distillation and has not established whether the original attacker ran the resale services. The report names OpenCode, Warp, Fireworks AI, and several others as disclosing similar thefts between October 6 and October 23. It cites Wiz’s “Black October” research as context only, and says that research does not attribute this incident.
How to check if your app is exposed
- Treat keys stored on Railway before October 21 as exposed unless Railway told you the workspace was outside the decrypted set. The 1.85% figure is an upper bound on environments. Revoke the provider key and issue a new one.
- Read the provider’s usage, not only Railway’s dashboard. Group spend by API key. Railway’s report says AWS Bedrock and Google have started fraud refunds, OpenAI wants a support case ID passed back to Railway, OpenRouter’s no-refund policy is one Railway says it will cover, and Railway will cover Anthropic charges while discussions continue. SpaceXSI claims go through Railway support.
- Confirm linked accounts. Railway says it reset BYOK agent keys and GitHub OAuth tokens for 1,852 users, notified each of them, and found no logins to Railway accounts using credentials taken in the incident. If you were in that set, the token you had connected is dead. Reconnect it yourself. Do not reuse a token that was present on October 21.
- Put the replacement in a server environment, then redeploy. A new key in
NEXT_PUBLIC_orVITE_is public. Railway restored builds by plan from 03:32 to 22:49 UTC on October 22. Running services kept running during the pause, so a revoked key could break production until a new deploy was possible.
How Sekrd helps
Sekrd does not read Railway’s secret store, and it would not have seen this decryption. It does catch the same class of key when a generated app also compiles it into JavaScript. A free scan flags OpenAI, Anthropic, xAI, Perplexity, and Fireworks key formats in the bundle, a Gemini key used against generativelanguage.googleapis.com, and server secret names behind NEXT_PUBLIC_, VITE_, EXPO_PUBLIC_, and the other client prefixes. Those rules are secret-pattern-match and js-client-secret-env.
If you are about to ship a replacement key from Cursor, v0, Lovable, Bolt, or Replit, scan the URL that users actually open. The steps for a key that already shipped are in the leaked-key guide and the env-prefix guide. A free scan is enough for the bundle. Pricing for a deep scan is on /pricing.
Related reading
- Chainalysis says live tokens for every user were exposed
- The October 8 tensorlake npm worm, which also hunted AI tool config
- Cursor security scanner and v0 security scanner
Frequently asked questions
Did Railway confirm that my API key was stolen?
No. The November 10 report says at most 1.85% of workspaces, about 100,000, could have had variables decrypted. Logs name environments, not individual keys, so the figure is an upper bound.
How much unauthorized AI spend did Railway report?
About $740,000. Railway says it will cover charges a provider will not refund. The report puts OpenRouter’s no-refund policy and ongoing Anthropic charges in that group.
Were customer databases or GitHub repos downloaded?
Railway says forensics did not show customer databases being accessed, and that customer repositories were not downloaded from the build system. It did read histories of 353 builds, including sealed variables, on October 20.
Is it safe to store a new key on Railway now?
The report says there was no customer-data access after 09:25 UTC on October 21 and that the weaknesses used are closed. Rotate the old key at the provider first. Sekrd does not audit Railway’s control plane.
Does this mean Lovable, Bolt, or Cursor were breached?
The Railway report does not say that. It does say stolen AI credentials later showed up in coding-agent traffic. A key compiled into your frontend is a separate problem a scan can see.
Sources
Don't ship until you're sekrd
Run a free scan to find the vulnerabilities your AI missed.
Scan Your App Free