The Agent Skills Directory

[PROMPT_INJECTION]: The generate_comparison_report function in references/EXAMPLES.md exhibits an indirect prompt injection vulnerability by processing untrusted data from an external tracking server without sanitization.\n
Ingestion points: Data is ingested from the MLflow tracking server via client.get_run(run_id) in references/EXAMPLES.md.\n
Boundary markers: No delimiters or instructions are provided to the agent or browser to ignore potential instructions embedded in the tracking data.\n
Capability inventory: The script possesses file-write capabilities (open(output_file, 'w')) to save generated reports locally.\n
Sanitization: The implementation is missing sanitization or escaping of the params and metrics values before they are rendered into the HTML output via pandas.to_html.\n- [COMMAND_EXECUTION]: The skill uses subprocess.check_output in references/EXAMPLES.md to retrieve git commit information. Although the command is currently limited to a static argument list (git rev-parse), this demonstrates a capability for executing shell-level commands.\n- [DATA_EXFILTRATION]: Documentation in references/EXAMPLES.md suggests configuring the MLflow server to listen on 0.0.0.0. This configuration exposes the tracking server to any network interface, potentially enabling unauthorized remote access to sensitive experiment metadata and artifacts.

track-ml-experiments