data-analysis

Pass

Audited by Gen Agent Trust Hub on Nov 16, 2026

Risk Level: SAFECOMMAND_EXECUTIONDATA_EXFILTRATIONEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTIONPRIVILEGE_ESCALATION
Full Analysis
  • [COMMAND_EXECUTION]: The skill relies on executing a Python script (analyze.py) that performs file system operations and manages a local DuckDB cache.
  • [DYNAMIC_EXECUTION]: The query action allows the agent to execute arbitrary SQL queries. This is a powerful feature but allows the execution of any valid DuckDB command, including functions that can access system resources.
  • [DATA_EXFILTRATION]: DuckDB's SQL engine contains functions such as read_text() and read_blob() that can access local files. If the agent is tricked into using these functions in the --sql parameter, it could expose sensitive files (e.g., .ssh/id_rsa, .env) that are accessible to the process.
  • [EXTERNAL_DOWNLOADS]: When processing .xlsx files, the script automatically executes INSTALL spatial; LOAD spatial; in DuckDB. This initiates an outbound network connection to DuckDB's official extension repository to download the spatial extension.
  • [PRIVILEGE_ESCALATION]: The --output-file parameter in analyze.py allows writing query results to any accessible path. Since the script creates directories and overwrites existing files without validation, it could be used to overwrite critical user files or configuration (e.g., shell profiles) if the path is maliciously constructed.
  • [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted user-supplied data (CSV, Excel) and incorporates it into a queryable database schema. It lacks explicit instructions to ignore embedded commands or boundary markers, making it potentially vulnerable to instructions hidden within the data cells that could influence the agent's analytical steps. Ingestion points: CSV and Excel files located in /mnt/user-data/uploads/. Boundary markers: None provided; the skill does not instruct the agent to ignore instructions embedded in the data. Capability inventory: File reading via SQL, file writing via the script's output parameter, and remote extension installation. Sanitization: SQL identifiers are sanitized, but cell content is not validated against instruction injection.
Audit Metadata
Risk Level
SAFE
Analyzed
Nov 16, 2026, 09:06 AM