The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: Unsafe file extraction in scripts/download_data.py. The extract_tar function uses tarfile.extractall() on a tar.gz archive downloaded from a remote URL. This method is susceptible to TarSlip/path traversal attacks, which could allow a malicious archive to write files to arbitrary locations on the host system if the source is compromised.
[COMMAND_EXECUTION]: Internal script execution via subprocess. scripts/data_client.py contains the update_data method which calls subprocess.run(['python', 'scripts/download_data.py', '--latest']). This pattern executes system commands to handle data updates.
[EXTERNAL_DOWNLOADS]: Fetching data from external third-party sources. The skill downloads datasets from the chenditc/investment_data repository on GitHub.
[PROMPT_INJECTION]: Potential surface for indirect prompt injection. (1) Ingestion points: scripts/download_data.py downloads data files from an external repository which are subsequently read by scripts/data_client.py. (2) Boundary markers: Absent. Data extracted from external sources is passed to the agent without delimiters or safety instructions. (3) Capability inventory: The skill includes subprocess.run() calls in scripts/data_client.py and file writing/extraction in scripts/download_data.py. (4) Sanitization: Absent. Data is loaded directly into pandas DataFrames and provided to the caller without filtering for malicious prompt instructions.

investment-data