The Agent Skills Directory

Command Execution: The skill executes gcloud and curl commands to manage cloud infrastructure, deployments, and lifecycle operations. While this is necessary for the skill's functionality, it involves direct interaction with the host environment and cloud resources.
Privilege Management: The instructions include steps for modifying IAM policy bindings, such as adding the Vertex AI Service Agent role to service accounts. This represents a security consideration as it involves managing access controls, though it is a required step for cross-project resource copying.
External Communication: The skill communicates with official Google Cloud APIs (googleapis.com) to deploy models and check statuses. These network operations target well-known, vendor-owned services related to the skill's primary purpose.
Indirect Prompt Injection Surface: The skill processes user-supplied prompts for testing deployments and ingests data from external model catalogs. The implementation of multi-tiered safety confirmation gates (Tier R, M, and D) serves as a review mechanism to mitigate these risks.
Ingestion points: The skill ingests data from the gcloud ai model-garden models list command and user-provided prompts for prediction testing.
Boundary markers: The skill explicitly defines "Safety & Confirmation Tiers" which require the agent to pause and wait for user approval before executing mutating or destructive actions.
Capability inventory: The skill utilizes the gcloud CLI for resource management, curl for API interactions using authentication tokens, and both bash and Python for utility scripts.
Sanitization: The instructions emphasize explicit user confirmation and cost-estimation cards as a primary control rather than automated data sanitization.
Dynamic Execution: The skill generates and executes local bash and Python scripts (e.g., calculate_cost.py) to automate tasks like price estimation and environment configuration.

agent-platform-deploy