The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/analyze.py is vulnerable to SQL injection. It constructs SQL queries by directly interpolating file paths and Excel sheet names into f-strings without escaping (e.g., in _load_excel and _load_csv). An attacker could provide a file with a crafted name (e.g., data.csv'); [SQL_PAYLOAD] --) to execute arbitrary SQL commands within DuckDB, which has built-in capabilities to read and write files.
[EXTERNAL_DOWNLOADS]: The script dynamically installs the duckdb and openpyxl Python packages from the official PyPI registry using pip at runtime if they are missing.
[EXTERNAL_DOWNLOADS]: The script uses DuckDB's INSTALL spatial; LOAD spatial; commands, which fetch the spatial extension from DuckDB's official remote servers during initialization.
[REMOTE_CODE_EXECUTION]: The script utilizes subprocess.run to execute package installation commands and interacts with the DuckDB shell engine. The SQL injection vulnerability provides a mechanism for indirect execution of commands through DuckDB's extended SQL functionality.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection through malicious file metadata.
Ingestion points: User-uploaded file paths in /mnt/user-data/uploads/ and internal Excel sheet names.
Boundary markers: None; data is directly interpolated into command strings.
Capability inventory: Execution of arbitrary SQL, file read/write access via DuckDB, and shell command execution for package management.
Sanitization: Partial regex-based sanitization is applied to table names, but the script fails to sanitize or escape the file paths and sheet names used in SQL queries.

data-analysis