AIDEFEND in Action AIA-0128 Incident

Railway: Build Isolation Failure and a Separate OIDC Re-entry Exposed AI Credentials

Railway is an application deployment platform. Railway reports an October 20 partial build-host filesystem escape, misuse of a build credential against Temporal and theft of its internal monorepo. After that route was contained, an erroneous OIDC binding enabled a second entry to production credentials and variable decryption. At most 1.85% of workspaces could have had variables decrypted; this is an upper bound, not proof that every variable was stolen. Subsequent misuse of AI keys was observed.

Credential TheftPrivilege EscalationRuntime IsolationAI Infrastructure
10 applicable AIDEFEND defenses
Source: Incident Report: October 21st Security Attack 
Author: Railway
Original article: Nov 10, 2026

Threat Analysis

  • Two access paths mattered: A build-host read exposed a credential accepted by Temporal. The attacker also obtained Railway’s monorepo, then found a separate OIDC trust error after the first route was contained.
  • Encryption did not remove authorization risk: Database reach and a decryption credential on an orchestration host together enabled reads of AI-related variables. The KMS root key itself remained inside KMS.
  • Impact is bounded by evidence: The workspace estimate is a maximum, not a count of fully stolen workspaces. Railway observed later AI-key misuse but did not establish that every subsequent user was the original attacker or that customer databases were accessed.

Applicable AIDEFEND Defenses (10)

AID-I-001.002
MicroVM & Low-Level Sandboxing
High
The hosting operator can run untrusted builds in dedicated disposable VMs with a separate host boundary, minimal mounts and isolated storage. Railway reports adding per-build ephemeral VMs. Verify that the actual builder cannot read host or neighboring-build sentinels and that the VM is destroyed after completion. The Framework’s Firecracker example is an implementation option, not an identified Railway component.
AID-H-004.002
Service & API Authentication
High
At the build-orchestration service boundary, bind each accepted machine credential to its intended service, workload and permitted API scope. Where implementing the Framework’s OAuth profile, verify issuer, audience, expiry and scope before the Temporal handler and block direct bypass. The report does not identify the old credential format, so this is a reviewed authentication integration, not a claim that the exploit was a particular JWT bug. Use independently scoped, expiring machine credentials, keep them out of durable build histories and deliver them only to the workload that needs them. Verify issuance, rotation and old-value rejection at the actual service. For decryption, bind each calling service to only its required secrets. Short lifetime reduces exposure duration but does not stop immediate misuse or an overly broad live credential. For affected customers using Amazon Bedrock, provider-side workload and endpoint policies can deny calls outside approved entry points, including runtime, batch and Mantle paths when enabled. Verify each real data-plane path and alternative role/key. This requires a suitable AWS workload/network deployment; it is not a Railway toggle and cannot be generalized to unrelated API-key providers.
AID-I-002.001
Internal AI Network Segmentation
High
Constrain the build plane’s internal routes and keep orchestration, secret storage and decryption on explicitly authorized paths. Test the effective reachability from actual build and orchestration workloads, including SSH and proxies. Do not grant database reach merely because a host deploys workloads.
AID-E-001.001
Root & Long-Lived Credential Object Eviction
High
Revoke exposed build, SSH, database, decryption and customer-provider credentials at their issuers, then deliver replacements only after the storage and execution environment is trustworthy. Verify every consumer and old-key rejection. Railway initially warned providers before requesting replacements because new keys in an uncleared platform could be stolen again.
AID-E-001.002
Issued Token, Authentication Session & Lease Revocation
High
Invalidate issued cloud, user and linked-account sessions as a separate response population. Where revocation is not available, deny the principal at the relying service and retain containment through expiry. Test actual reuse; resetting a local environment variable does not revoke a token held elsewhere.
AID-E-001.003
AI Agent & Workload Principal and Issuance Disablement
High
Disable the exact compromised cloud federation binding and reconcile its IaC source before allowing CI to resume. Independently read back issuer, subject, audience and effective trust policy, then prove the formerly accepted workload cannot mint a fresh credential. The provider adapter must match the actual cloud; the public report does not identify it.
AID-R-006
Clean-Room AI System Redeployment Orchestration
High
Restore builders, internal tooling, deployment identities and secret-decryption services through an explicit recovery dependency graph. Pin each clean artifact and recovery workflow, and require separately verified isolation and federation fixes before dependent components receive new credentials or CI authority. Railway’s decision to rebuild internal tooling with CI disabled illustrates this ordering requirement. Independently read back the assembled deployment and probe both access routes before cutover.
AID-D-011.004
Non-Human Identity & Delegated Token Abuse Detection
Medium
Join workload credential issuance, federation and resource use against an approved service/resource baseline. Correlate a build identity reaching history APIs, unusual cloud issuance, bulk variable reads and decryption fan-out. Validate telemetry completeness first; encrypted storage alone does not make an authorized-looking decrypt safe. Affected AI workload owners should inspect provider-side invocations independently of Railway or an application proxy. For Bedrock, collect the enabled model data events and applicable native findings, then reconcile authorized direct-use exceptions and suspicious workload/model/region or volume changes. Other providers require their own verified audit adapter; missing logs cannot be called a clean result.
AID-R-003.003
Progressive Capability & Privilege Re-enablement
Medium
Restore CI, builders and decryption privileges in independently verified stages, with limited canaries and explicit rollback. Re-enable only after the original and secondary routes fail safe tests and replacement credentials are confined. A service appearing healthy is not evidence that its old trust relationship is safe.
AID-R-004
Fleet Remediation Propagation & Technical Recurrence Prevention
Medium
Require all builders and deployment identities to inherit the verified isolation, service-authentication and secret-service policies. Gate IaC changes and reconcile running instances, including legacy systems. Railway’s separate recovery architecture is evidence of an applicable operator-owned boundary, not proof that every example in this Framework was deployed.

What Defenders Should Do Now

  • Platform operators should isolate untrusted builds, authenticate internal services for their exact purpose, validate federation trust and separate secret-reading from broad deployment authority. The unpublished exploit details remain unknown.
  • Contain each identity layer separately. Disable the federation path, revoke exposed long-lived credentials and issued sessions, and verify authoritative rejection before distributing replacements into rebuilt infrastructure.
  • Customers should inspect provider-side usage and rotate affected keys at the actual issuers. Where supported, provider-native workload restrictions can constrain direct misuse; a local gateway policy alone cannot. Restore capabilities progressively and enforce the verified baseline across every builder.

Conclusion

The central lesson is that closing one runtime escape does not close an independent identity path. AIDEFEND  connects AI-credential protection to build isolation, service trust, decryption authorization, provider-side detection and verified recovery. Each control must cover its own deployment boundary.

Last updated: