AIDEFEND in Action AIA-0120 Incident

Zeabur October Breach: AI Credentials Exposed Across Cloud and Cluster Boundaries

Zeabur, a platform for hosting applications and databases, reported an October 14-15 intrusion affecting hosted AI-service credentials. Attackers used leaked internal credentials across cluster and cloud boundaries; a replacement credential was captured in a later backup and subsequently used without authorization. Customers own upstream-key revocation and recovery of deployments they administer; the operator owns internal isolation and backup access.

Credential ExposureCredential GovernanceRuntime IsolationAI Infrastructure
9 applicable AIDEFEND defenses
Source: Zeabur October 2026 security incident (preliminary report) 
Author: Yuanlin Lin
Original article: Oct 19, 2026

Threat Analysis

  • Two paths, one observed convergence. An already-held administrator token reached the public cluster interface and service secrets. An old-backup route to a secrets-service key is assessed; subsequent production-secret reads are confirmed. Separately, an old GCP key could read plaintext root passwords from initialization settings, an assessed explanation for confirmed root logins. The attacker combined a compromised server with internal-network credentials to enroll a Tailscale node, then queried the platform database with an over-broad read-only account. Initial credential acquisition is unknown.
  • AI credentials were a concrete target. Container commands targeted new-api/one-api, gateways that manage upstream AI access, and n8n, a workflow automation platform that stores third-party integration credentials. Some commands failed, and outputs were not recorded. Do not equate attempted reads with a complete credential-by-credential exfiltration inventory.
  • A replacement can be exposed again. A newly rotated credential was captured in a later cluster backup and subsequently used without authorization. The backup retrieval path is assessed. Containment and backup access must be addressed alongside rotation.
  • Keep confirmed effects separate from hypotheses. The link to August is unconfirmed. Amazon Bedrock is AWS’s hosted model service. Reported Bedrock calls all failed without charges; attribution of a user’s later OpenAI usage to this leak is also a hypothesis. The report does not establish prompt injection or AI-agent execution of the intrusion.
  • Responsibility follows the boundary. Customers can revoke their issuer credentials, review their own usage and recover systems they administer. The operator owns cluster administration, internal-network admission, production-data access and platform backups. Moving secrets out of the platform database does not protect them on a compromised host.

Applicable AIDEFEND Defenses (9)

AID-E-001.001
Root & Long-Lived Credential Object Eviction
High
Customers should revoke exposed upstream AI/API keys at their issuers, then replace them only in clean, approved consumers. Operators own internal cloud, database, signing and enrollment credentials. Enumerate exact key IDs and consumers, retain provider readback, and verify that the old credential is denied while the clean replacement performs only its intended task. The framework’s executable IAM/Secrets Manager/ECS example needs a different native adapter for Kubernetes, GCP, SSH or another issuer; it is not a ready-made Zeabur script. First contain ongoing readers and backup access so replacement secrets are not exposed again. Already issued sessions and copied data require separate handling.
AID-I-002.001
Internal AI Network Segmentation
High
The platform operator should allow production-database connections only from explicitly approved services, not every node carrying a broad internal-network tag. Separate customer workloads from internal services and enumerate indirect VPN, proxy, shared-identity and cloud-control paths. Use Tailscale policy and cloud/network enforcement at boundaries the compromised workload cannot edit. Independently read effective rules and test allowed service access plus denial from an unapproved node. Kubernetes NetworkPolicy needs an enforcing CNI and is not a reliable barrier against an attacker holding cluster-administrator authority; the production boundary needs independent administration.
AID-I-003.001
Detection-Driven Quarantine & Safe-Mode Downgrade
High
After a verified incident finding, the operator can quarantine the exact compromised workload or server, block its data-access routes and retain only an approved forensic channel. Bind the action to authenticated incident evidence, target identity and current policy; use native cloud or network controls outside the compromised host. Preserve evidence when safe, read back the applied state and test that the target cannot reach protected services. Credential revocation, deleting an enrolled node and trusted rebuilding remain separate actions. A customer cannot quarantine Zeabur’s internal cluster through an application setting.
AID-E-001.002
Issued Token, Authentication Session & Lease Revocation
High
Invalidate already-issued platform login sessions and access tokens at every verifier, separately from replacing the signing key. For internal-network access, enumerate and remove the unauthorized enrolled node and verify its connection is denied; revoking an enrollment credential alone is insufficient. Tailscale documents this distinction for auth keys, but Zeabur does not disclose the exact credential type used to enroll the node. Use provider-native session/device APIs rather than the framework’s Keycloak example, reconcile all applicable artifact classes and test old-artifact denial plus clean authorized access. The effect is stopping established access, not recovering disclosed data.
AID-E-001.003
AI Agent & Workload Principal and Issuance Disablement
High
For the exposed old GCP service-account key, the operator can disable its exact compromised service account through the authoritative IAM control plane and stop fresh credential issuance, rather than deleting only one key. Resolve the immutable service-account identity and every issuance path; separately confirm the production credential’s type and owner instead of assuming it shares this identity. Use a prepared native adapter, preserve readback and test fresh authentication or minting denial. Provision a separately trusted least-privilege replacement if needed. Google documents that key deletion does not revoke already-issued short-lived credentials; their end-to-end denial remains the issued-access response. Disabling a shared account affects legitimate consumers and needs an approved containment/recovery plan.
AID-D-005.001
AI System Log Generation & Collection
Medium
Operators should collect cluster secret reads and exec requests, cloud object reads and setting writes, network enrollment and database access with authenticated identity, resource, result and UTC time. Send available records to a separately administered retention-locked archive; reconcile coverage with controlled canary operations and test protected-version deletion denial. Enable the source audit coverage first: archiving cannot recreate reads that were never logged. Keep raw API keys, passwords and database contents out of ordinary telemetry. Customers can preserve their issuer and server records, but need the operator for internal platform evidence. This supports detection and investigation; it does not itself identify or block an attack.
AID-R-006
Clean-Room AI System Redeployment Orchestration
Medium
For an affected AI deployment with several dependent components, the owner should rebuild in a clean environment using an exact recovery manifest: approved application/image digests, database recovery point, newly provisioned credentials and the order for reconnecting services. Use trusted workflow templates and separately verify live component identities and external readiness before cutover. Restore reviewed data rather than blindly cloning a root-compromised disk or reusing its configuration. This needs a prepared deployment/recovery pipeline and verified backups. A single password reset, a server reinstall command or an unchecked backup is not evidence that the integrated deployment is clean.
AID-H-004.002
Service & API Authentication
Medium
At an AWS account the defender administers, replace unnecessary static integration keys with supported short-lived workload credentials and restrict Bedrock models, regions and invocation routes through provider IAM. The workload must not be able to edit its own boundary or assume an unbounded alternate role. Inventory public endpoints and enabled runtime, batch and other invocation paths; test permitted access and provider-authorized denial outside the approved route. This constrains follow-on use of an exposed AWS credential. Zeabur reported failed Bedrock attempts, not successful paid model use, and did not establish why each request failed. Other AI providers require their own supported identity and enforcement mechanisms.
AID-D-011.004
Non-Human Identity & Delegated Token Abuse Detection
Partial
For the reported Bedrock-attempt branch, the AWS account owner can correlate native invocation events, credential identity and approved workload routes, retaining denied attempts and direct calls that have no gateway match. Enable the applicable CloudTrail data events and reconcile accounts, regions and missing partitions; management-event history alone is insufficient. Validate the detector with authorized fixtures covering a legitimate direct call, unauthorized use and missing telemetry, then check delivery to responders. This detects attempted or actual provider-side credential abuse within the measured population. It does not establish that the original secret read was detected, that model use succeeded or that any GuardDuty finding occurred in this incident.

What Defenders Should Do Now

  • Follow Zeabur’s all-user guidance without waiting for an email: inventory still-valid credentials stored on the platform, including replacements. Revoke at the issuer, update clean approved consumers and verify old-credential denial. Use secure channels and credential IDs rather than posting secret values.
  • For affected n8n deployments, plan replacement of the encryption key together with re-creation of third-party credentials; test recovery before reconnecting workflows. For new-api/one-api, revoke upstream AI keys and regenerate gateway access tokens. Replacing a database connection environment variable does not change the database’s actual password.
  • Treat identified root-compromised GCP servers as untrusted. Preserve evidence when safe, isolate them, rebuild from trusted sources and restore reviewed data. Inventory both cloud-managed SSH settings and local authorized keys. A root-password reset does not remove SSH keys or prove persistence is gone.
  • Platform responders should remove unauthorized enrolled nodes, revoke their enrollment authority and invalidate affected sessions separately. Fence old backup access before placing replacement secrets into recoverable backups. Verify production routes from actual approved and unapproved workloads.
  • Review issuer-side access, model usage and billing plus available server and application records. Ask Zeabur for case-specific internal evidence and its coverage limits. Keep absent or incomplete logs as unknown, rather than evidence of safety.
  • Prepare a verified migration and data-export plan. Zeabur’s October 28 follow-up confirms planned December 31 closure and says it cannot pay third-party incident charges under its current finances; it offers assistance with waiver or refund requests.

1 additional consideration

Database authorization and secret-bearing backups

A permitted private route still needs narrowly scoped database authorization. Backup readers and decryption authority must be separated from compromised workloads; encryption at rest does not protect plaintext from an identity permitted to decrypt it.
Recommendation: Operators should restrict database roles and query scope, remove unnecessary secret copies and separate backup access authority. Customers should request evidence for the operator-owned boundary and verify recoverability before migrating.

Conclusion

A credential rotation is durable only when the attacker can no longer read the replacement or keep using access already established. AIDEFEND  maps this incident to issuer-side credential eviction, independent internal segmentation, incident-driven quarantine, issued-access invalidation, protected evidence and verified recovery. Provider-side authentication and identity-abuse monitoring further constrain and expose attempted downstream model use. Customers and operators must execute the controls they actually own, then verify the relevant access boundary before restoring service.

Last updated: