Skip to main content
Back
IntrusionKubernetes · Google Cloud · AWS · SaaS
14 min read

Agent-driven intrusion from an exposed Kafka UI across Kubernetes, Google Cloud, AWS and the identity provider

Victim
An unnamed large multi-cloud organization investigated by the Wiz Customer Incident Response Team
When
not stated; the post gives no date for any stepDisclosed Oct 19, 2026
Entry
A Java remote code execution flaw in an internet-exposed Kafka UI instance, one of the confirmed initial access vectors per Wiz; no CVE or version is named, and the other vectors are unconfirmed.

Wiz CIRT describes one intrusion at a large multi-cloud organization that began with a Java remote code execution flaw in an internet-exposed Kafka UI and gives no date for any step. From the pods the actor took service account tokens, environment secrets and node credentials, then minted Google Cloud credentials for dozens of identities, scraped registries, deployed public Cloud Run services, ran AWS Systems Manager commands, assumed OrganizationAccountAccessRole with a new IAM user and created an identity provider administrator that signed in to AWS until it was suspended. Through CI/CD runners and a GitOps controller it reached cluster-admin and dumped Secrets across clusters; RBAC held for Secrets in some cases and bucket creation was denied, while the outcome of the database, vault and exfiltration attempts is not stated, and that the actor was an AI agent is Wiz's reading of the behavior.

What happened

External Parties
Attacker-controlled hosts
Public request catcher
Internet
Cloud APIspod, node and minted credentials
Kubernetes clusters and CI/CD
AWS Organization
+ token-creator service account
+ registries scraped in a loop
+ public pentest- Cloud Run
Google Cloud
+ administrator API key
+ pentest administrator user
+ groups mapped to AWS roles
Identity Provider Tenant
Kafka UI
Application Pods
Secrets and ConfigMaps
GitOps Controller
CI/CD Runners
Systems Manager Instances
probe IAM User
Member Account Admin Role
Databases
IAM Identity Center

Replaying the 2026 Kafka UI agent intrusion / no step date published; Wiz CIRT case of October 19, 2026

  1. 1

    Exposed Kafka UI

    Date not stated

    A Java flaw in an internet-exposed Kafka UI runs shell commands, and command-and-control agents spread to pods across several clusters.

  2. 2

    Credentials from the pods

    Date not stated

    Mounted service account tokens and secret-like environment variables are read in the pods, and node credentials come from the instance metadata service.

  3. 3

    Google Cloud and AWS

    Date not stated

    A token-creator service account mints credentials in Google Cloud, and in AWS Systems Manager commands and a new probe IAM user reach administrative access across member accounts.

  4. 4

    Identity provider backdoor

    Date not stated

    A compromised identity provider API key creates a pentest administrator, which signs in to AWS through IAM Identity Center within minutes.

  5. 5

    Pipelines, clusters and exfiltration

    Date not stated

    CI/CD runners and a GitOps-synced manifest give cluster-admin, Secrets are dumped across clusters, and data leaves through several routes.

Impact

From an exposed Kafka UI the actor reached pods across several clusters and took service account tokens, environment secrets and node cloud credentials. It minted Google Cloud credentials for dozens of identities, scraped container registries, deployed public Cloud Run services, ran commands through AWS Systems Manager, reached AWS databases over SSH and gained administrative access across AWS member accounts. An identity provider administrator it created signed in to AWS until the account was suspended, and CI/CD and GitOps paths gave it cluster-admin and Secrets across clusters. RBAC held for Secrets in some cases and the bucket creation attempts were denied. Which exfiltration routes delivered data, and how much, is not stated.

Outcome

Where the provider's controls fit

Five controls would have refused a step; a card names the condition where one applies. The combined inbound rules of all attached security groups on AWS admitting the UI port only from the organization's own ranges and excluding the attack source, or a hierarchical firewall policy on Google Cloud would have refused the exploit request to the Kafka UI, a management console that need not serve the internet. IMDSv2 with a hop limit of 1 would have refused an initial node-credential read from an unprivileged pod on the AWS pod network while the actor could neither launch a host-network pod nor change the node. Wiz does not establish whether the first metadata read preceded the GitOps-derived cluster-admin privilege, which provides a bypass; later harvesting by that identity is not prevented by this setting. A service control policy would have refused the creation of the probe user in a member account, and a resource control policy that refuses IAM users on OrganizationAccountAccessRole would have refused its step into your member accounts. Every other cloud control here only bounds a step: egress rules on the nodes and the runners, security group references on the databases, the GKE metadata server, an IAM deny policy on token minting, organization policy constraints and VPC Service Controls in Google Cloud, and service control policies on Systems Manager and on role sessions from unknown networks. Wiz does not say which identities made those calls or from where, so how far each would have held is not known. No cloud control reaches the Kafka UI flaw, the pods' own environment, the identity provider tenant, the CI/CD platform, the GitOps controller or Kubernetes RBAC; Kubernetes admission policies, network policies and service account token settings bound parts of those steps.

Could this have been prevented?

In hindsight, not as a whole. 5 controls would have refused a step outright, and 22 controls would have limited one. 9 steps have no cloud-provider control. The counts below are control assessments, not attack stages: several can address the same step. The same replay with the controls in place before the incident shows where each phase would have been refused or bounded.

5
Prevented
22
Contained
9
Not covered

How it could have been prevented

The same replay with the controls in place. Green marks where a step is refused, blue where it is bounded, and red where no control reaches it. Open a row to read the control, why it acts, and its caveat.

External Parties
Attacker-controlled hosts
Public request catcher
Internet
Cloud APIspod, node and minted credentials
Kubernetes clusters and CI/CD
AWS Organization
+ token-creator service account
+ registries scraped in a loop
+ public pentest- Cloud Run
Google Cloud
+ administrator API key
+ pentest administrator user
+ groups mapped to AWS roles
Identity Provider Tenant
Kafka UI
Application Pods
Secrets and ConfigMaps
GitOps Controller
CI/CD Runners
Systems Manager Instances
probe IAM User
Member Account Admin Role
Databases
IAM Identity Center

Replaying the 2026 Kafka UI agent intrusion / no step date published; Wiz CIRT case of October 19, 2026

1 · Exposed Kafka UI

2 Prevented4 Contained1 Not covered4 on other clouds

2 · Credentials from the pods

1 Prevented4 Contained1 Not covered2 on other clouds

3 · Google Cloud and AWS

2 Prevented6 Contained1 Not covered12 on other clouds

4 · Identity provider backdoor

1 Contained1 Not covered2 on other clouds

5 · Pipelines, clusters and exfiltration

7 Contained5 Not covered2 on other clouds

Key facts

  • Wiz CIRT published the case on October 19, 2026, and gave no date or time for any step; when the intrusion started, how long it ran and when it was contained are not stated
  • Before the entry the actor probed several internet-facing services with exploit payloads, anonymous cloud storage requests and API abuse of an exposed internal developer tool, with no stated outcome; the same source IP scanned more than 600 other Wiz customer environments during the same window, per Wiz on October 19, 2026
  • A Java remote code execution flaw in an internet-exposed Kafka UI instance, with no CVE or version named, ran shell commands from the Java process; malware, a command-and-control channel and agents in pods across several clusters followed, and internal message brokers, schema registries and databases were scanned to map reach
  • In Google Cloud a service account provisioned with token-creation privileges generated credentials for dozens of identities, container registries were scraped in a continuous loop and publicly accessible pentest- Cloud Run services were deployed; the outcome of an attempted exfiltration of high-value database records is not stated
  • In AWS, Systems Manager ran commands across instances and a new IAM user prefixed probe assumed OrganizationAccountAccessRole for administrative access across member accounts; a pentest administrator created with a compromised identity provider API key signed in to AWS through IAM Identity Center within minutes and kept sessions until it was suspended
  • Five CI/CD workstreams ran in parallel, a GitOps-synced manifest escalated to cluster-admin and Secrets were dumped across clusters; RBAC blocked Secrets in some cases and attempts to create task-named buckets were denied, while the outcome of the attempts against the secrets vault, and which exfiltration routes delivered data, are not stated

Sources

Every fact in this breakdown comes from the public reporting below. The replay is a model of that reporting, not a reconstruction of the environment it describes.

  1. 1.
    Loud, Fast, and Everywhere: Anatomy of a Real AI-Driven Cloud AttackWiz CIRT · Oct 19, 2026 · ResearchEntry vector, credential theft in workloads, Google Cloud token minting, AWS Systems Manager and role assumption, identity provider backdoor user and single sign-on, CI/CD and GitOps escalation, Kubernetes Secret access, exfiltration channels, denied actions

The estate in the replay is illustrative and drawn at the level of attack stages; the victim's real environment is not public. Verdicts describe what a control does when it is in place before the incident. They are not an assessment of the victim's security program.

More cloud attack breakdowns

All breakdowns