• Sources: primary, discussion, discussion
  • Summary: Railway's 2026-11-10 report says an attacker partially escaped filesystem isolation in its multi-tenant build infrastructure on 2026-10-20, took a build credential that its internal Temporal cluster accepted as a trusted client, and read build histories and a copy of Railway's monorepo. Using a misconfigured OIDC binding found in that code, the attacker obtained cloud credentials, reached an orchestration machine holding a long-lived decryption credential, and decrypted variables for at most 1.85% of workspaces, about 100,000, querying specifically for AI provider keys. Railway estimates unauthorized spend on stolen keys at about $740,000, reset linked tokens for 1,852 users, and is moving builds to per-build virtual machines.
  • Why it matters: Any platform that stores customer secrets should check whether internal credentials are accepted beyond their purpose and whether CI OIDC trust bindings are scoped tightly.
  • Background: Envelope encryption encrypts each secret with a data key and then encrypts that data key with a root key held in a key management service, so stealing the database alone yields only ciphertext. The weak point becomes any machine allowed to ask the KMS to decrypt, which is what the attacker reached here. Railway says OpenCode, Warp and Fireworks AI disclosed similar thefts between 2026-10-06 and 2026-10-23, and it treats this as one of a series of attacks aimed at platforms that hold AI keys.
  • Follow-up: Refunds of the roughly $740,000 in unauthorized AI spend from providers or Railway.
  • Comments: The Hacker News thread drew 20 points and 2 comments, with no added detail.