← Tất cả tin tức

Các bài viết được đăng bằng tiếng Anh.

Threat IntelligenceAI SecurityGovernance AI

Attackers Decrypted Railway Customers' Environment Variables and Went Straight for AI Provider Keys, Running Up About $740,000. Could One of Your Keys Be Spent Without You Noticing?

Obiguard Research Team·19 tháng 11, 2026·4 min read

Railway, the cloud deployment platform, says an attacker who escaped a build VM in October decrypted customers' environment variables, prioritised rows that looked like AI provider keys, and caused an estimated $740,000 in unauthorized usage. The account comes from Railway's own incident report, updated on 10 November 2026, and was covered by The Hacker News. We read Railway's report; we have not independently verified its claims, and we could not open The Hacker News page.

What Railway reports

  • Initial access (20 October, UTC): an unknown vulnerability partly escaped filesystem isolation on a multi-tenant build VM. The resulting build credential was trusted by an internal workflow system, letting the attacker read stored histories of 353 builds, including sealed variables.
  • Escalation (21 October): an erroneously defined OIDC binding gave the attacker a short-lived cloud credential, which led to internal credentials and a machine holding a long-lived decryption key.
  • What was taken: the attacker queried encrypted environment-variable rows, prioritising names matching AI provider keys, and decrypted some. Railway says at most 1.85% of workspaces, roughly 100,000, could have had variables decrypted, and calls that an upper bound. Linked-account tokens for 1,852 users were also in the decrypted data and were reset.
  • Not found: Railway says forensics did not indicate customer databases were accessed.
  • Eviction: access ended at about 09:25 UTC on 21 October.
  • The cost: Railway says unauthorized spend on the stolen keys was estimated to reach about $740,000. It does not explain how it estimated that figure.
  • Why AI keys: Railway believes the keys were meant for resale through relay services. It says it has not established whether the attacker operated those services.

On refunds, Railway says AWS Bedrock and Google Gemini are refunding under existing fraud policies, OpenAI cases are reviewed individually, and Railway is covering costs itself where providers will not pay, including OpenRouter. Its position on refunds shifted during the incident, so check the current report.

Why this matters beyond Railway

A stolen database password lets an attacker read data. A stolen AI provider key lets them spend your money, immediately, with no exploit left to run. The key is bearer-style: whoever holds it is you, as far as the provider is concerned.

We have made a related argument about stolen AI API keys used as cover. The Railway incident shows the other half: even when you store keys in a platform's encrypted variables, a compromise of the platform can expose them. Rotation after the fact is necessary but slow, and the spend happens in the gap.

What to check now

  1. Inventory every AI provider key, including ones in deployment platforms, CI settings and no-code tools. Many teams cannot list them.
  2. Rotate keys held on any platform that has disclosed an incident, then review provider usage from the incident window for spend you cannot explain.
  3. Set hard spend limits and alerts at each provider, and per project where supported.
  4. Issue one key per application or agent, so a leak can be revoked without breaking everything and usage can be attributed.
  5. Restrict which models and tools each key can reach, rather than leaving every key with access to everything.
  6. Prefer short-lived credentials over long-lived static keys where the provider allows it.

How Obiguard Governance AI fits

Steps 4 and 5 are the part a key alone cannot do. Governance AI puts your AI traffic behind one gateway, so applications and agents call models with Obiguard credentials instead of holding raw provider keys everywhere. Its allow-lists limit which models and tools a given agent may use, and every call is recorded in the audit ledger under a named agent, so unusual usage is visible and a single credential can be revoked.

Scope matters. Governance AI governs traffic that goes through it. A provider key that is still valid and already stolen can be used directly against the provider, so the provider key behind the gateway must be stored and rotated carefully, and keys that never moved behind the gateway stay exposed. To see how it would fit your setup, get in touch or read about Governance AI.

Sources

Obiguard giúp bạn thế nào

Biến điều này thành chính sách được thực thi, không chỉ là nhận thức.

Obiguard nằm phía trước mọi yêu cầu AI mà tổ chức của bạn thực hiện — sàng lọc prompt và đầu ra theo các rào chắn bảo vệ, khung tuân thủ và chuỗi kiểm toán mà những câu chuyện như thế này khiến trở nên cần thiết.

Xem cách hoạt động →