customer-support-analysis

Pass

Audited by Gen Agent Trust Hub on Sep 3, 2026

Risk Level: SAFECOMMAND_EXECUTIONDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTIONDATA_EXFILTRATION
Full Analysis
  • [COMMAND_EXECUTION]: The script asa.py invokes warehouse command-line tools such as bq, snow, and databricks via subprocess.run to execute SQL queries. This is the primary function of the skill and is performed using the user's existing CLI configurations.
  • [DYNAMIC_EXECUTION]: The generate-dashboard.py script facilitates the creation of interactive dashboards by embedding agent-generated JavaScript code into an HTML template. This code uses the Chart.js library to render visualizations based on query results.
  • [INDIRECT_PROMPT_INJECTION]: The skill processes external data sourced from Zendesk tickets. This introduces a theoretical surface for indirect prompt injection where malicious content within a ticket could attempt to influence the agent's behavior during analysis.
  • Ingestion points: Data is ingested through SQL queries against Zendesk models in BigQuery, Snowflake, or Databricks (e.g., SKILL.md mentions zendesk__ticket_enriched).
  • Boundary markers: The instructions do not define specific adversarial delimiters for separating query results from the rest of the prompt context.
  • Capability inventory: The agent possesses the capability to execute shell commands (bash) and Python scripts (python3) to perform data analysis and dashboard generation.
  • Sanitization: Content rendered in the HTML dashboard is generally escaped using html.escape in generate-dashboard.py, although the system supports raw HTML for specific UI elements like status badges.
  • [DATA_EXFILTRATION]: asa.py performs network requests to api.fivetran.com to retrieve configuration metadata. These operations are consistent with the skill's authorship by Fivetran and are necessary for resolving data source locations.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 3, 2026, 09:23 PM