Vertex AI Agent Builder — agentic threat model

Vertex AI hosts foundation models (including pre-trained models) making it susceptible to adversarial prompt injection, model extraction/stealing, and membership inference attacks targeting proprietary training data.

With AutoML integration and custom training routines, the platform is highly exposed to data poisoning, training data exfiltration, and lineage/provenance gaps during dataset ingestion and preprocessing.

Not certain from the listing — while it is an 'Agent Builder', the description focuses heavily on ML workflows rather than specific runtime orchestration frameworks, leaving potential gaps in tool-calling security and memory poisoning controls.

As a cloud-hosted platform, deployment infrastructure risks include container escape, privilege escalation within Google Cloud projects, and unauthorized access to model endpoints.

The platform explicitly supports robust monitoring and versioning, which helps mitigate model drift, but remains vulnerable to evaluation gaming and sophisticated adversarial evasion that bypasses standard monitoring.

Features built-in governance and compliance controls aligned with enterprise standards, but misconfigurations in IAM policies or tenant isolation could lead to unauthorized model modification.

Not certain from the listing — the description does not specify if the platform supports multi-agent collaboration or a shared agent marketplace, which would introduce risks of cascading failures and agent-to-agent trust abuse.