sales-pipeline-analysis

Pass

Audited by Gen Agent Trust Hub on Sep 3, 2026

Risk Level: SAFECOMMAND_EXECUTIONEXTERNAL_DOWNLOADSDYNAMIC_EXECUTIONDYNAMIC_CONTEXT_INJECTIONINDIRECT_PROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill uses asa.py to interact with various data warehouse CLI tools (bq, snow, databricks) via the subprocess.run method to execute queries and check authentication status.
  • [EXTERNAL_DOWNLOADS]: The asa.py script communicates with api.fivetran.com using the urllib.request module to fetch account information, connector configurations, and transformation metadata.
  • [DYNAMIC_EXECUTION]: The skill employs dynamic execution patterns for utility purposes:
  • In SKILL.md, python3 is invoked via a heredoc (<<'PYEOF') to format query results into Unicode box-drawing tables.
  • The asa.py script contains a MOCK_FETCHER hook that can execute an arbitrary binary specified in the ASA_FIVETRAN_FETCHER environment variable, intended for testing environments.
  • The generate-dashboard.py script takes a JSON payload and injects a string from the script field directly into an HTML <script> tag to render Chart.js visualizations.
  • [DYNAMIC_CONTEXT_INJECTION]: The SKILL.md file suggests the use of the !command syntax (specifically ! gcloud auth application-default login) as a shortcut for users to authenticate their local CLI environment within the agent's context.
  • [INDIRECT_PROMPT_INJECTION]: The skill is designed to process external data retrieved from HubSpot CRM (deals, stages, and history). While this presents a potential surface for indirect prompt injection if data contains malicious instructions, the skill implements defensive measures:
  • Ingestion points: Data enters the context from BigQuery, Snowflake, or Databricks via asa.py resolution.
  • Boundary markers: The instructions explicitly tell the agent to "State facts" and "Never assert what you can't see in the data."
  • Capability inventory: The skill can execute warehouse CLI commands and open local HTML files.
  • Sanitization: Labels are sanitized using TRIM() in SQL queries, and the dashboard generator uses HTML escaping for data values.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 3, 2026, 09:23 PM