datahub-evals
Pass
Audited by Gen Agent Trust Hub on Oct 1, 2026
Risk Level: SAFECOMMAND_EXECUTIONEXTERNAL_DOWNLOADSCREDENTIALS_UNSAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill makes extensive use of the Bash tool to run the acryl-datahub-cloud CLI for listing, fetching, and reporting evaluations. It also instructs the agent to use claude --print to initiate subagent runs.
- [EXTERNAL_DOWNLOADS]: Instructions guide the installation of the acryl-datahub-cloud package (specifically the datahub-evals extra) using pip install. This tool is the core component provided by the service vendor for performing evaluations.
- [CREDENTIALS_UNSAFE]: The skill utilizes DataHub connection credentials, explicitly mentioning the use of the ~/.datahubenv configuration file and environment variables such as DATAHUB_GMS_URL and DATAHUB_GMS_TOKEN to authenticate the CLI.
- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted evaluation questions fetched from DataHub and passes them as prompts to subagents.
- Ingestion points: Evaluation questions are fetched from the DataHub instance via acryl-datahub-cloud evals get as described in SKILL.md.
- Boundary markers: The command templates in SKILL.md do not implement specific prompt delimiters for the interpolated evaluation questions, though the skill suggests using system prompt instructions to guide subagent behavior.
- Capability inventory: Subagents are typically configured with MCP tools (e.g., DataHub server) to interact with the catalog; the primary agent has access to Bash and Task tools as defined in SKILL.md.
- Sanitization: The skill advises mitigating risks by narrowing the tool surface using --strict-mcp-config and --allowedTools to restrict the subagent's capabilities to only the necessary DataHub MCP server.
Audit Metadata