vertex-agent-builder
Pass
Audited by Gen Agent Trust Hub on Sep 9, 2026
Risk Level: SAFECOMMAND_EXECUTIONEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill uses the
Bashtool to executegcloudCLI commands for project configuration, agent deployment, and health monitoring. These operations are within the scope of the skill's primary purpose and target well-known Google Cloud administrative interfaces. - [EXTERNAL_DOWNLOADS]: Instructions include the installation of official Google Cloud libraries (e.g.,
google-cloud-aiplatform,vertexai) from established package registries. These dependencies are provided by a well-known service and are required for the skill's functionality. - [INDIRECT_PROMPT_INJECTION]: The skill provides scaffolds for agents that process untrusted external data through RAG pipelines and multimodal extraction.
- Ingestion points: Untrusted data enters the agent context via GCS buckets, BigQuery sources, and PDF/image uploads as detailed in
references/examples.mdandPRD.md. - Boundary markers: The generated scaffolds include system instructions and "citation-first" policies intended to ground agent responses in retrieved context.
- Capability inventory: The skill environment grants access to file system operations and shell execution tools.
- Sanitization: The implementation patterns rely on standard library parsing and user-defined validation checklists to handle external inputs.
Audit Metadata