What the demonstration did
Replaying Unit 42's ModeLeak demonstration / published November 2024; the run dates are not stated
- 1
A custom job runs as the service agent
Not dated · before November 12, 2024With only aiplatform.customJobs.create, a custom job runs a researcher-built image as the Custom Code Service Agent, and its metadata server hands out that agent's credentials and the VM's user-data script.
- 2
Source project data and internal images
Not dated · before November 12, 2024The service agent's token lists, reads and exports the source project's buckets and BigQuery datasets, and the user-data script leads to images in internal Google repositories.
- 3
A poisoned model in the prediction tenant
Not dated · before November 12, 2024A model re-imported with a swapped container image starts a shell when deployed, and the read-only prediction service account lists its permissions and the tenant's GKE clusters.
- 4
Model images and adapters copied
Not dated · before November 12, 2024Cluster credentials give read-only kubectl access in two tenant clusters, the images of other deployed models are pulled and exported, and the caip buckets with the tuned LLMs' adapter files are copied.
In Unit 42's test environment, one custom job permission reached the source project's buckets, BigQuery datasets and service account list and images in internal Google repositories; a poisoned model reached the prediction tenant's clusters, the images of every deployed model and the tuned LLMs' adapter files. Kubernetes writes and Docker use were refused.
What ModeLeak shows about your managed ML identities
One permission to create a custom job was enough, because the job ran as a Google-managed service agent with broad rights in your project. On your Google Cloud projects, a custom constraint that requires a service account on custom jobs keeps them off that service agent, and an IAM deny policy on the project's service agents and a VPC Service Controls perimeter bound what its token reads. An IAM deny policy on model upload and deploy narrows who can place a model in serving. No customer setting reaches the metadata server, the clusters, the repositories or the buckets of Google's tenant projects, where most of the chain ran. Google implemented fixes there, and no source says what it changed or when.
Would the controls have stopped it?
In hindsight, no control would have stopped it. 6 controls would have limited how far it went. 6 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.
- 0
- Prevented
- 6
- Contained
- 6
- 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 Unit 42's ModeLeak demonstration / published November 2024; the run dates are not stated
1 · A custom job runs as the service agent
2 Contained1 Not covered3 on other clouds2 · Source project data and internal images
2 Contained1 Not covered3 on other clouds3 · A poisoned model in the prediction tenant
2 Contained1 Not covered3 on other clouds4 · Model images and adapters copied
3 Not coveredKey facts
- One permission, aiplatform.customJobs.create, was enough to start chain 1: the custom job ran a researcher-built image as the AI Platform Custom Code Service Agent and opened a reverse shell, per Unit 42
- The job's metadata server returned the service agent's credentials and the VM's user-data script, which exposed VM creation details and metadata about internal Google Artifact Registry repositories
- With the service agent's token, the researchers listed, read and exported Cloud Storage buckets and BigQuery datasets their own identity could not reach, and downloaded images from internal Google repositories; the authors could not fully measure that last impact
- A model exported from the Model Registry, with the container image in its environment.json swapped, was re-imported and deployed, and started a shell in the prediction tenant project as the custom-online-prediction service account, which the authors describe as read-only
- Unit 42's write-up of November 12, 2024 reported 3 GKE clusters in the tenant project and read-only kubectl access to 2 of them; the researchers pulled the images of other deployed models and copied caip buckets holding the tuned LLMs' adapter files, whose usability was not shown
- Kubernetes create, delete, update, attach and exec attempts failed, and Docker could not run; Unit 42 says Google implemented fixes for both issues, and no fix date, CVE or Google statement is published
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.ModeLeak: Privilege Escalation to LLM Model Exfiltration in Vertex AIUnit 42 (Palo Alto Networks) · Nov 12, 2024 · ResearchCustom job escalation chain, poisoned model chain, service agent and service account identities, permissions used, data and models reached, what held, vendor fix, recommendations
- 2.Google Vertex AI Vulnerability Exposes Customer DataeWeek · Nov 27, 2024 · NewsRe-report of both flaws and the patch claim
The estate in the replay follows the researchers' description of their own test environment. Verdicts describe what a control does when it is in place before the run. They are not an assessment of any company's security program.