GCP Vertex AI High GenerateContent Volume From Single IP

IMPORTANT: This documentation is no longer updated. Refer to Elastic's version policy and the latest documentation.

GCP Vertex AI High GenerateContent Volume From Single IP

Detects a single source IP making a high number of Vertex AI GenerateContent (or StreamGenerateContent) audit calls against the same model resource in the lookback window. That pattern fits model extraction, automated scraping, or a compromised caller burning prediction quota.

Rule type: esql

Rule indices: None

Severity: medium

Risk score: 47

Runs every: 10m

Searches indices from: now-60m (Date Math format, see also Additional look-back time)

Maximum alerts per execution: 100

References:

Tags:

Domain: GenAI
Domain: Cloud
Data Source: GCP Vertex AI
Data Source: GCP
Data Source: Google Cloud Platform
Platform: GCP
Service: GCP Vertex AI
Use Case: Model Theft
Use Case: Threat Detection
Tactic: Collection
Threat: LLMjacking
Threat: Unauthorized AI Usage
Mitre Atlas: AML.T0024
Mitre Atlas: AML.T0024.002
Mitre Atlas: AML.T0029
Resources: Investigation Guide
Rule Type: ES|QL

Version: 1

Rule authors:

Elastic

Rule license: Elastic License v2

Investigation guide

edit

Triage and analysis

Investigating GCP Vertex AI High GenerateContent Volume From Single IP

A single source.ip issued an unusually high number of Vertex AI GenerateContent / StreamGenerateContent audit calls against one model resource (gcp.vertexai.audit.resource_name) in the lookback window. That pattern is consistent with model extraction, automated scraping, or a compromised caller burning prediction quota.

Possible investigation steps

Note source.ip, gcp.vertexai.audit.resource_name, cloud.project.id, and Esql.event_count on the alert. Compare the count to the project’s normal peak for that model.
Confirm whether the IP is known egress for an approved app, CI runner, or shared NAT. Check Esql.event_action_values for StreamGenerateContent vs GenerateContent mix.
Pivot to logs-gcp_vertexai.prompt_response_logs-* for the same time window and model: sample prompts, token volume, and whether content looks like systematic extraction (repetitive probes, large context dumps).
Pivot auditlogs for the same IP: other aiplatform methods, new service accounts, or SetPublisherModelConfig changes that could hide or redirect logging.
Identify client.user.email on matching GenerateContent audit events to map IP → principal.

False positive analysis

Shared corporate NAT or proxy egress aggregating many legitimate users onto one IP.
Approved batch evaluation, load tests, or data-migration jobs. Validate with change tickets and expected schedules before escalating.

Response and remediation

If unauthorized: revoke or rotate credentials for the calling principal, restrict the IP via VPC Service Controls / firewall, and apply per-IP or per-principal quotas.
Review billing and quota usage for the model; hunt for similar volume from other IPs in the same project.

Setup

edit

Setup

Requires GCP Vertex AI auditlogs for aiplatform.googleapis.com. Raise Esql.event_count above your normal peak before broad enablement.

Rule query

edit

from logs-gcp_vertexai.auditlogs-*
| where
    data_stream.dataset == "gcp_vertexai.auditlogs" and
    source.ip is not null and
    (
        event.action like "*PredictionService.GenerateContent*" or
        event.action like "*PredictionService.StreamGenerateContent*"
    ) and
    (gcp.vertexai.audit.status.code is null or gcp.vertexai.audit.status.code == 0)
| stats
    Esql.event_count = count(*),
    Esql.timestamp_first_seen = min(@timestamp),
    Esql.timestamp_last_seen = max(@timestamp),
    Esql.event_action_values = values(event.action)
  by
    source.ip,
    gcp.vertexai.audit.resource_name,
    cloud.project.id
| where Esql.event_count >= 100
| keep
    source.ip,
    gcp.vertexai.audit.resource_name,
    cloud.project.id,
    Esql.event_count,
    Esql.timestamp_first_seen,
    Esql.timestamp_last_seen,
    Esql.event_action_values

Framework: MITRE ATT&CK^TM

Tactic:
- Name: Collection
- ID: TA0009
- Reference URL: https://attack.mitre.org/tactics/TA0009/
Technique:
- Name: Data from Cloud Storage
- ID: T1530
- Reference URL: https://attack.mitre.org/techniques/T1530/

« Entra ID Privileged Role Assigned to Another Account GCP Vertex AI Publisher Model Config Modified »