What was exposed
Replaying the GKE Sys:All January 2024 exposure
- 1
Configured: broad rights for every Google account
Date not stated; before the January 2024 disclosureThe company's GKE clusters bind broad Kubernetes rights, stated as cluster admin, to system:authenticated, a group that in GKE holds any Google account.
- 2
Reachable: any Google account reads the ConfigMaps
Not statedA generic Google token lists and reads the company's ConfigMaps, where a shell script holds an AWS access key with broad S3 rights.
- 3
Reachable: the AWS key reads several S3 buckets
Not statedThe key from the ConfigMap lists and downloads several S3 buckets, whose logs hold administrator credentials for internal systems.
- 4
Found and closed: Orca reports, Google blocks
Bulletin on Jan 19, 2024; other dates not statedOrca discloses to the company, which tightens its IAM roles, secures its S3 buckets and removes the secrets from its ConfigMaps; Google publishes GCP-2024-003.
Any Google account could read the company's ConfigMaps, and through the AWS key in them several S3 buckets whose logs held administrator credentials for internal systems; how long this stood is not stated, and no source says that anyone other than Orca used the access.
Where provider and platform controls fit
The binding that gave every Google account broad rights is the step a control refuses: the GKE setting that disallows bindings to system:authenticated would have kept the group out, although it came after the incident and does not remove a binding that already exists. A Kubernetes ValidatingAdmissionPolicy rejects such a binding only while the policy stands, and a principal that can bind cluster admin can remove it. Control plane authorized networks would have capped who reaches the cluster's IP-based endpoints to your listed ranges. Where your AWS account is a member account of an AWS Organization, a service control policy or a resource control policy with network conditions would have capped reads of the S3 buckets to your approved networks; Azure, Google Cloud and Oracle Cloud bound the same read with their own network rules. The key in a ConfigMap, the credentials in your logs and your response to the finder's report stay outside the provider's controls.
Could this have been prevented?
In hindsight, not as a whole. 1 control would have refused a step outright, and 5 controls would have limited one. 3 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.
- 1
- Prevented
- 5
- Contained
- 3
- 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.
Replaying the GKE Sys:All January 2024 exposure
1 · Configured: broad rights for every Google account
1 Prevented1 Contained2 · Reachable: any Google account reads the ConfigMaps
1 Contained1 Not covered3 · Reachable: the AWS key reads several S3 buckets
3 Contained1 Not covered2 on other clouds4 · Found and closed: Orca reports, Google blocks
1 Not coveredKey facts
- The company's GKE clusters bound broad rights, stated as cluster admin, to system:authenticated, a group that in GKE holds any Google account; the overview of the method speaks of custom roles instead
- Orca authenticated with a generic Google token that any Google user can obtain, listed and read the company's ConfigMaps, and listed and pulled images from its container registries
- A shell script in a ConfigMap held an AWS access key and secret with broad S3 rights; with it Orca listed and downloaded the contents of several S3 buckets
- The downloaded logs held administrator credentials for internal systems, among them a platform used by the company's customers, a message broker, a search cluster and an authentication server; the write-up did not say they were used
- Orca's census found about 1,300 potentially vulnerable GKE clusters of more than 250,000 scanned, 108 of them immediately compromisable
- Google published bulletin GCP-2024-003 on January 19, 2024, notified affected customers and blocked new cluster-admin bindings to the system groups from GKE 1.28; no source said whether anyone other than Orca used the access
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.How the Sys:All Loophole Allowed Us To Penetrate GKE Clusters in ProductionOrca Security · Jan 24, 2024 · ResearchThe group binding, any Google account as the reader, the ConfigMap with the AWS key, the S3 rights, the listing and download of S3 buckets, the administrator credentials in the logs, the registry image pulls, the company's fixes
- 2.Sys:All: GKE system group exposure (Orca Research Pod)Orca Security · Jan 24, 2024 · ResearchThe system:authenticated group in GKE, the census of about 250,000 clusters, about 1,300 potentially vulnerable and 108 immediately compromisable, Google's block of cluster-admin bindings in GKE 1.28, Google's view that the behavior is intended
- 3.GKE security bulletins: GCP-2024-003Google Cloud · Jan 26, 2024 · AdvisoryBulletin dates, severity medium, clusters with bindings to system:authenticated, GKE 1.28 blocks new cluster-admin bindings to system groups, Google's recommendations
- 4.'Significant security loophole' found in Google software container systemThe Record by Recorded Future News · Jan 24, 2024 · NewsGoogle's statement, the bulletin sent to affected users, the company's AWS credentials as reported
The estate in the replay is illustrative and drawn at the level of the exposure; the owner's real environment is not public. Verdicts describe what a control does when it is in place before the data is exposed. They are not an assessment of the owner's security program.