The Agent Skills Directory

Command Execution Considerations: The skill documents usage of gcloud and curl for interacting with Google Cloud APIs. This process involves generating bearer tokens (gcloud auth print-access-token) and transmitting them in request headers. Users should manage these tokens carefully to avoid accidental disclosure.
Indirect Prompt Injection Surface: The skill describes the creation of agent containers that ingest system_instruction and data from external sources like Google Cloud Storage (GCS). This constitutes an injection surface where external content could influence agent behavior.
Ingestion points: The system_instruction field and mounted directories (GCS or Skill Registry).
Boundary markers: No specific delimiters or instructions to ignore embedded commands are shown in the configuration examples.
Capability inventory: Agents can be equipped with code_execution, filesystem, google_search, and url_context tools.
Sanitization: The skill does not define methods for sanitizing the ingested instructions or remote data before they are processed by the agent.
Agent Capability and MCP Security: The skill enables agents with execution and network capabilities. It includes guidance on configuring Model Context Protocol (MCP) servers, highlighting that the platform is designed to maintain header confidentiality during tool execution.

gemini-agents-api