News · 11 min read

Railway Hack and Zeabur Breach: Your AI Keys May Be Exposed

On October 21, 2026, Railway, a hosting service (a company that runs other businesses' apps), warned its users about the Railway hack. Someone may have read the secret settings customers keep there, some in readable form. A week earlier, Zeabur, another hosting service, had a break-in, and its attacker went looking for AI keys.

We read both reports, Railway's recovery guides and two more October disclosures to show what each company confirmed.

Below, we cover what happened, why API keys (secret codes that let software use an account) were the target, and how to fix it.

Executive Summary

  • Railway says some stored secrets may have been read in plain text, so it asks every user to replace them, just to be safe.
  • Zeabur confirms that 3,218,965 stored settings and data from 321,810 accounts were read, and it is closing down by December 31, 2026.
  • The fix is the same for both: turn off each old key with the company that issued it, make a new one, update your services and check your bills.
  • OVHcloud and PingCAP had separate leaks the same month, through different routes, and no company has tied all four to one attacker.
  • For document work, Elephas is a more private choice than ChatGPT, Gemini and Claude, as it hides personal details on your Mac first and its on-computer model needs no AI key.

The Railway Hack Puts Every Stored Secret at Risk

Railway's October 21 disclosure says an outsider may have reached the system that holds customers' environment variables (settings and secrets passed to running software), and decrypted (turned back into readable text) some of them.

Our card quoting Railway's own words from its October 21 security disclosure
Our card quoting Railway's own words from its October 21 security disclosure

The disclosure says the harm may be limited to a subset of users. Even so, it tells everyone to replace every stored variable, every API key and any private registry login (the login used to download private software packages). Railway will email the customers it confirms were affected.

Railway believes this is part of a series of attacks on platform companies, aimed at AI inference providers (the services that run AI models and answer requests). It says it told those providers to cancel keys, but it did not name them. When we checked on October 22, there was no full incident review yet.

Railway also paused new deployments (putting new versions of software live) while it worked. Our timeline below puts the main steps from Railway's status page in order (all times UTC, the world standard time).

TimeWhat Railway posted
October 21, 21:59Security notice linked; deployments paused for every plan
October 21, 22:10All storage bucket keys being reset
October 22, 03:32Builds back for Enterprise customers
October 22, 04:17All Railway API tokens cancelled
October 22, 06:29Builds back for Pro customers
October 22, 13:53Pro and Enterprise deployments at normal speed; Hobby, Free and Trial still waiting

The Zeabur Hack Exposed Millions of Stored Settings

Zeabur's incident report confirms that an attacker read 3,218,965 stored settings and data linked to 321,810 accounts during a break-in on October 14 and 15.

Our breakdown of the record, account, key and token counts in Zeabur's report
Our breakdown of the record, account, key and token counts in Zeabur's report

Zeabur takes the blame in plain words. Long-lived admin logins (master logins that never expired) were stored in plain text and could be used from the open internet. Zeabur says it cannot confirm how the attacker first got them.

The report says the attacker was mainly hunting for keys they could use straight away, most of all keys for AI services. The attacker also tried to use Zeabur's own Amazon keys on Amazon Bedrock (Amazon's service for AI models). Every one of those tries failed, and no charges were made.

Zeabur is also shutting down. Some services start closing on December 1, 2026, and all dedicated servers (computers kept for one customer) and shared clusters (groups of computers that host many customers) close by December 31.

Two More Key Leaks Hit the Same Month

OVHcloud, a cloud hosting company, and PingCAP, which runs the TiDB Cloud database service, both disclosed leaks on October 19 or 20, and each one came through a different door.

Our side-by-side view of the four incidents, how each happened and what was confirmed
Our side-by-side view of the four incidents, how each happened and what was confirmed

The OVHcloud advisory says someone used a developer's GitHub account to swap the files of its CLI (a tool you control by typing commands) for harmful ones. Only versions v0.11.0 to v0.15.0 downloaded on October 19, between 14:47 and 16:21 Paris time, were affected. Version v0.16.0 is clean.

PingCAP's update says stolen service-account logins (logins used by software, not a person) gave an outsider access from September 9 to October 7. The copied data included a locked table of TiDB Cloud API keys. One tested key could not be used, and investigators found no sign that data stored in TiDB was touched.

A group called FulcrumSec claims it took far more from PingCAP. Dark Reading says it did not confirm those claims on its own, so we treat them as claims only.

IncidentWhenHow it happenedWhat was confirmed
RailwayOctober 21, 2026An outsider reached the system that holds customer secretsSome secrets may have been read in readable form; all users told to replace keys
ZeaburOctober 14 to 15, 2026Leaked admin logins kept in plain text3,218,965 stored settings and 321,810 accounts read
OVHcloudOctober 19, 2026A developer's GitHub account swapped download filesVersions v0.11.0 to v0.15.0 in a short window were harmful
PingCAPSeptember 9 to October 7, 2026Stolen service-account loginsData copied, including a locked table of API keys

AI Keys Are the Prize Because They Spend Your Money

Railway's AI guides tell users to look for requests and spending they do not recognise, and Zeabur says its attacker was mainly hunting for AI keys.

Our table of what each kind of stolen key can let someone do
Our table of what each kind of stolen key can let someone do

An AI key works like a prepaid card for AI. Whoever holds it can send requests that are billed to your account. That is why Railway's OpenAI and Anthropic guides ask you to check usage by key and compare the models, requests and costs with your normal use.

Other keys open other doors. A storage key can open your files, and a database password can open the data that account can reach. Railway's bucket guide asks you to look for files you do not recognise or files that are missing.

We wrote about a similar case earlier this month. The OpenCode breach also exposed AI keys that people had saved in a hosted tool, though nobody has linked it to these attacks.

Fix It by Cancelling Each Key Where It Was Made

Railway and Zeabur both ask you to replace exposed secrets, and Zeabur says to cancel the old ones with the company that issued them first.

Our step-by-step checklist for each platform
Our step-by-step checklist for each platform

Start with a list of every key and password you kept on the platform. Give each one an owner on your team, and write down when the old one stopped working. This is our suggestion for keeping the work in order.

Railway's credential reset guide lists what Railway already reset and what you must still do. It reset bucket keys and GitHub tokens, so keys from companies such as OpenAI are yours to replace. Update your settings, redeploy when your plan can build again, reconnect GitHub, and cancel old registry tokens on GitHub.

Check your bills and usage while you work, and save any odd records for the provider's support team. Our guide on local vs cloud AI explains why where your AI runs matters too.

Elephas Keeps Your Documents Stored on Your Mac, Not on Hosted Servers

Every tool that keeps your keys or files on servers you cannot see is one more place a break-in can reach, and the same is true of the AI tools you use for work.

Our view of what leaves your Mac on each of Elephas's three answer routes
Our view of what leaves your Mac on each of Elephas's three answer routes

With ChatGPT, Gemini and Claude, your questions and uploaded files go to their servers. Depending on your plan and settings, chats can be kept and may be used to train their models. Elephas is private AI for your documents, and you choose where each answer is made.

With Railway and Zeabur, Stored AI Keys Became the Target

With Elephas, the on-computer model (an AI model that runs on your Mac) answers without sending your questions or documents anywhere, and it needs no AI key at all. There is no key sitting on a server for anyone to take.

With Railway and Zeabur, Secrets Sat on Hosted Servers

With Elephas, your documents stay stored on your Mac. When you do use a cloud model, Elephas hides names, email addresses and phone numbers on your Mac before the question leaves, then puts them back in the answer. You can open a short note under each answer to see what was hidden.

With Elephas AI, your question and the passages it picked go through the Elephas gateway (the service that passes the request on) to open-source models in Elephas's own cloud. They are not stored and are not used to train AI models.

With Railway and Zeabur, Every Stored Key Had to Be Replaced

With Elephas, the Team plan gives everyone one shared, managed setup. Staff do not need to paste their own AI keys into different tools, and the owner decides whether members can use on-computer models.

Conclusion

Private AI for your documents, on your Mac

Elephas answers from the files on your Mac, and with its on-computer model your questions and documents are not sent anywhere. On cloud routes it hides names, emails and phone numbers on your computer first. Plans start at $19 a month, and Team is $29 per person per month.

Start the 7-day free trial
Selvam Sivakumar
Written by

Selvam Sivakumar

Founder, Elephas.app

Selvam Sivakumar is the founder of Elephas and an expert in AI, Mac apps, and productivity tools. He writes about practical ways professionals can use AI to work smarter while keeping their data private.

← Back to Resources