lakehouse-explorer
Pass
Audited by Gen Agent Trust Hub on Aug 6, 2026
Risk Level: SAFECOMMAND_EXECUTIONEXTERNAL_DOWNLOADSDATA_EXFILTRATIONPROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill uses
bash,python3, andduckdbto perform administrative and analytical tasks. It follows security best practices by implementing an 'EXPLAIN gate' that requires explicit user confirmation of resource usage and query plans before executing potentially costly operations. - [EXTERNAL_DOWNLOADS]: The skill downloads DuckDB core extensions (
httpfs,iceberg) from official DuckDB repositories and optionally installs thepyicebergpackage from the Python Package Index (PyPI). These are well-known, trusted resources required for the skill's primary functionality. - [DATA_EXFILTRATION]: The skill handles sensitive credentials (client secrets) by reading them from a local
.lakehouse_config.jsonfile. It follows secure practices by only transmitting these credentials to the user-specifiedpolaris_endpointand caching tokens locally in/tmpusing standard permissions. It does not send data to unauthorized third-party domains. - [PROMPT_INJECTION]: The skill is theoretically susceptible to indirect prompt injection via malicious metadata in the remote Polaris catalog or Iceberg tables. However, it mitigates this risk by utilizing a three-tier decision process that prioritizes metadata-only lookups and enforces hard caps on query results (LIMIT 100).
- Ingestion points: Data enters the agent's context through Polaris REST API responses and Apache Iceberg table data/metadata retrieved via DuckDB (SKILL.md).
- Boundary markers: The instructions mandate a multi-step 'EXPLAIN gate' to visualize and confirm operations before data processing.
- Capability inventory: The skill utilizes shell execution, Python script generation, and SQL querying capabilities (SKILL.md).
- Sanitization: Implements strict query rules including mandatory filters, aggregations over raw data retrieval, and hard limits on returned rows.
Audit Metadata