If your apps run on Railway, replace the keys stored there and check your AI bills. Whatever your host, give every AI key a spending limit and an owner.
- 01Railway said on 10 November 2026 that an October break-in could have exposed stored keys in at most 1.85% of workspaces, which Railway puts at about 100,000 workspaces.
- 02Railway’s investigation suggests AI keys were the main target. It estimates unauthorised spend on the stolen keys at about $740,000.
- 03OpenAI and Anthropic both document monthly spending limits; Anthropic says setting them through its software interface is in early access. A limit caps most of the bill (Noesis analysis).
- 04This week: list where each AI key lives, who owns it and what it may spend.
In plain words
- Hosting service
- A company that runs your website or app on its computers, such as Railway. It often also stores the passwords and keys the app needs.
- AI key
- A secret code that lets an app use an AI service such as OpenAI or Anthropic. Whoever holds it can run up charges on the owner’s account.
- Workspace
- Railway’s name for one customer’s account area, holding that customer’s apps and stored keys.
- Spending limit
- A cap on what an AI account can spend each month, set with the provider. OpenAI says requests fail once spend reaches the cap, though spend can run slightly over.
Railway says attackers could read stored keys in up to 1.85% of customer workspaces
On 10 November 2026 Railway published its final report on a break-in at its service. Railway hosts websites and apps for other businesses. It also stores the passwords and keys those apps use. Railway says an attacker broke in on 20 October and was shut out by 09:25 UTC on 21 October. The attacker could have read stored keys in at most 1.85% of workspaces, which Railway puts at about 100,000 workspaces. A workspace is one customer’s account area. Railway calls that figure an upper bound. Its investigation suggests AI keys were the primary target. Railway estimates unauthorised spend on the stolen keys at about $740,000. See source 1. Noesis analysis: spend on a stolen key is billed to the account that owns it.
From break-in to final report
- Attacker signs upFirst attacker accounts created
- Break-inStarted 07:28 UTC; Railway aware at 13:58 UTC
- Attacker shut outBy 09:25 UTC
- Preliminary reportCustomers who joined before 21 October asked to replace stored keys
- Final reportUp to 1.85% of workspaces; spend on stolen keys estimated at about $740,000
This reaches any business that keeps AI keys with a hosting service
If your website, app or internal tool runs on Railway, this is directly about you. Railway says it emailed the customers it judged at high risk. It says wider notice will follow after a set period, but gives no date. See source 1. Noesis analysis: no email from Railway does not prove you were unaffected. The lesson also reaches businesses on other hosts. An app that uses an AI service usually holds a key for it, often stored with the host. If an agency built an AI feature for you, ask where its keys live.
A stolen AI key works like a company card in a stranger’s hands
An AI key lets an app use an AI service, and the account that owns the key pays. Railway’s database logs show the attacker searching for AI keys first. Railway asked GitHub, GitLab, Anthropic, OpenAI, OpenRouter, SpaceXSI, Google and Stripe to cancel keys that may have been taken. See source 1. Noesis analysis: treat each AI key like a company card. It needs a limit, a named holder and one job, agreed between your head of technology and finance lead. The comparison stops at refunds, which here depend on each host and provider.
Who holds what for an AI key
Hosting service
- Stores the key next to your app
- Protects its own systems
- Tells customers after a break-in
AI provider
- Charges the account that owns the key
- Can cancel keys
- Offers spending limits
Your business
- Knows where each key lives
- Sets the limit and the alert
- Checks the bill each month
The trap is an AI key nobody owns, with no spending limit
Railway says it will cover breached-key costs with Anthropic for now, and all breach costs with OpenRouter. For OpenAI, Railway says affected customers should open a support request with OpenAI and pass the case ID to Railway support. Railway asks all customers to report charges they did not make. See source 1. Noesis analysis: the hard case is a key a past developer or agency set up, billed to a card nobody watches, with no limit. OpenAI lets an account set a hard monthly limit for the whole organisation or for one project. Once spend reaches it, requests fail, though spend can run slightly over first. See source 3. Anthropic’s documentation describes monthly limits for an organisation and for each workspace, and says setting them through its software interface is in early access. See source 4. Noesis recommendation: a limit caps most of the bill but also stops the AI feature, so set it with headroom and an alert.
Questions for whoever runs your apps, this week
Send these to your developer, agency or head of technology, and ask for written answers within the week. Noesis recommendation: name one owner for AI keys, such as the head of technology, and copy finance on every limit. The figure below shows what a good answer looks like.
- Which of our apps use an AI service, and where is each key stored?
- Do any of them run on Railway? If so, have we replaced the keys stored there, as Railway asked of customers who joined before 21 October?
- Does each AI account have a monthly spending limit, and who gets the alert?
- Does each key do one job, so we can cancel it without stopping everything?
- Who checks AI bills each month, and who acts on a charge we did not make?
What a good answer looks like
- 01Where is each AI key?Pass when: A written list of apps, keys and where each is stored
- 02Anything on Railway?Pass when: Keys stored there replaced, with the date
- 03Is there a spending limit?Pass when: A monthly limit and a named person on the alert
- 04One job per key?Pass when: A key can be cancelled without stopping other apps
- 05Who checks the bill?Pass when: A monthly check, and a route for disputed charges
What Railway’s reports do not tell you
Railway calls 1.85% an upper bound, because its logs show groups of stored settings rather than single keys. It has not said which customers were affected beyond those it emailed. It says it has no sign that the attacker downloaded customers’ code. It reset the keys 1,852 users had linked to other services, and found no logins to Railway accounts with them. These figures are Railway’s own, read on 13 November 2026. Disclosure: Noesis hosts this website on Railway. We have not seen any customer’s records, and we name no business as affected. See sources 1 and 2.
Railway’s own figures
- Upper limit on the share of workspaces whose stored keys could have been readClaimed by Railway
- 1.85%
- Estimated unauthorised spend on stolen keysClaimed by Railway
- $740,000
- Users whose keys linked to other services were resetClaimed by Railway
- 1,852
For your technical team.
Forward this section. Railway’s final report (10 November 2026) says the attacker partially escaped filesystem isolation in the build system at 07:28 UTC on 20 October and read the stored histories of 353 builds, sealed variables included. On 21 October an erroneously defined OIDC binding gave the attacker a short-lived cloud credential, and a subset of environment variables was decrypted. Railway disabled active cloud credentials and revoked static ones at 07:14 UTC on 21 October, and invalidated all user sessions and tokens at 20:19 UTC. Linked-account tokens (BYOK agent keys and GitHub OAuth tokens) for 1,852 users were decrypted and reset. Railway’s fixes: a dedicated ephemeral VM per build, and central audit logging of every encrypt and decrypt operation. The preliminary report (22 October) asked customers who signed up before 21 October to rotate keys stored on Railway, remove public network access to databases, check GitHub audit logs and app access, and check registry audit logs such as Docker Hub. OpenAI: organization and project hard limits return HTTP 429 with organization_spend_limit_exceeded or project_spend_limit_exceeded, and enforcement is not instant. Anthropic: Claude Console organizations set organization and workspace spend limits; setting them through the Admin API is in early access. Noesis recommendation: issue one key per app and environment, alert on usage, and keep a register of where each key is stored. See sources 1–4.
Sources & context
Sources checked 2026-11-13. Recommendations and illustrative scenarios are Noesis’s analysis, not claims of client results.
- 1 / Railway — Incident report: October security incident; published 10 November 2026; read 13 November 2026 ↗
- 2 / Railway — Preliminary report: October security incident; published 22 October 2026; read 13 November 2026 ↗
- 3 / OpenAI — Spend limits (API documentation); page date not stated; read 13 November 2026 ↗
- 4 / Anthropic — Set spend limit (Claude Platform documentation); page date not stated; read 13 November 2026 ↗
