The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests data from a local SQLite database and incorporates it into generated Markdown reports, which are subsequently read by the agent. This creates a surface where malicious data in the database could influence agent behavior.\n
Ingestion points: Data fields such as username, description, and payload are retrieved from the protocol and command_log tables in puda.db via extractor.py.\n
Boundary markers: No delimiters or instructions are used in report.py to distinguish untrusted data from the rest of the report content.\n
Capability inventory: The skill possesses capabilities to write files to the local file system, including Markdown reports, PNG visualizations, and CSV/JSON exports.\n
Sanitization: The ExperimentReport class in report.py performs no escaping or validation of database-sourced strings before they are interpolated into the Markdown templates.\n- [DYNAMIC_EXECUTION]: The skill employs a pluggable registry architecture that dynamically executes functions based on measurement types.\n
Evidence: In plotter.py and report.py, plotting functions are registered in a dictionary and invoked at runtime. Specifically, the add_plot method in ExperimentReport executes a provided callable with arbitrary keyword arguments (plot_function(**plot_kwargs)). While this is a standard design for extensibility, it constitutes dynamic execution of local code paths.

puda-data