store-performance-analysis

Pass

Audited by Gen Agent Trust Hub on Sep 3, 2026

Risk Level: SAFECOMMAND_EXECUTIONCREDENTIALS_UNSAFEINDIRECT_PROMPT_INJECTIONPRIVILEGE_ESCALATIONDYNAMIC_EXECUTION
Full Analysis
  • [COMMAND_EXECUTION]: The asa.py script makes multiple calls to subprocess.run to execute warehouse CLI tools such as bq, snow, and databricks. These calls are used to perform database queries and verify authentication states, which is central to the skill's functionality.
  • [CREDENTIALS_UNSAFE]: The skill interacts with sensitive local configuration and credential files. Specifically, it reads from ~/.databrickscfg to resolve Databricks host information and manages its own credential store at ~/.fivetran/credentials.json where it persists Fivetran API tokens.
  • [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process data from external Shopify tables, including customer names, product titles, and refund notes.
  • Ingestion points: Data is pulled from Shopify tables (e.g., customer, product, order) via warehouse CLIs into the agent's context.
  • Boundary markers: The instructions do not define explicit boundary markers or "ignore instructions" delimiters for the data returned by these queries.
  • Capability inventory: The skill has access to the bash tool, including python3 and various database CLIs.
  • Sanitization: There is no evidence of sanitization or filtering of the external data before it is presented to the agent.
  • [PRIVILEGE_ESCALATION]: The instructions include a "Codex Databricks Override" logic that directs the agent to attempt rerunning commands "outside the sandbox or with elevated permissions" if specific error codes are returned by the Databricks CLI. While intended to resolve environment-specific credential caching issues, this represents a guided sandbox escape pattern.
  • [DYNAMIC_EXECUTION]: The asa.py script contains a developer/testing hook via the ASA_FIVETRAN_FETCHER environment variable. If set, the script will execute the provided command string to perform HTTP requests instead of using standard libraries, allowing for dynamic command injection if the environment is compromised.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 3, 2026, 09:15 PM