adk-deploy-guide

Pass

Audited by Gen Agent Trust Hub on Mar 5, 2026

Risk Level: SAFEPROMPT_INJECTIONEXTERNAL_DOWNLOADSCOMMAND_EXECUTION
Full Analysis
  • [PROMPT_INJECTION]: The guide describes architectural patterns for event-driven agent invocations in references/event-driven.md. These patterns involve processing external untrusted data (Pub/Sub, Eventarc, BigQuery) as agent input, which presents a surface for indirect prompt injection.\n
  • Ingestion points: references/event-driven.md defines FastAPI endpoints for Pub/Sub, Eventarc, and BigQuery Remote Functions.\n
  • Boundary markers: The provided Python code examples do not include explicit delimiters or 'ignore' instructions for the incoming message payloads.\n
  • Capability inventory: The skill leverages high-privilege capabilities including infrastructure deployment via terraform apply, cloud resource management through gcloud, and CI/CD operations via gh auth.\n
  • Sanitization: The implementation patterns demonstrate direct injection of external payloads into the agent runner without explicit sanitization or validation logic.\n- [EXTERNAL_DOWNLOADS]: The skill references documentation and configuration files from google.github.io (a well-known service) and utilizes the uvx tool to execute the agent-starter-pack utility.\n- [COMMAND_EXECUTION]: The documentation provides instructions for executing essential deployment commands such as make deploy, adk deploy, terraform apply, and gh auth login as part of the standard production workflow.
Audit Metadata
Risk Level
SAFE
Analyzed
Mar 5, 2026, 05:36 PM