The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes user-supplied disease names to automatically generate report filenames. This creates a vulnerability surface for path traversal.
Ingestion points: The disease_name variable originates from user input as described in the 'Core Workflow' section of SKILL.md.
Boundary markers: There are no explicit validation checks or boundary markers to ensure the input is a valid disease name before it is used in file operations.
Capability inventory: The skill utilizes file-writing capabilities (open(filename, 'w')) to save research data.
Sanitization: The provided Python snippet only replaces spaces with underscores (replace(' ', '_')), failing to filter for directory traversal sequences like ../ or ..\.
[COMMAND_EXECUTION]: The instructions command the agent to perform local file system operations. Specifically, the create_report_file and update_report_section functions in SKILL.md perform file creation and modification based on user-controlled strings, which could lead to unauthorized file manipulation.
[EXTERNAL_DOWNLOADS]: The skill integrates with a vast array of external 'ToolUniverse' tools (e.g., OSL_get_efo_id_by_disease_name, OpenTargets_get_associated_targets_by_disease_efoId, PubMed_search_articles). This extensive reliance on external API calls for data retrieval represents a significant attack surface for potential data exfiltration if the tools or the agent's context are manipulated.
[DYNAMIC_EXECUTION]: The skill includes explicit Python code blocks for the 'Research Protocol' and instructs the agent to use this logic for its workflow. This encourages the agent to generate and execute script logic at runtime to manage the research report lifecycle.

tooluniverse-disease-research