The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests data from external, third-party sources including patent databases (via patent.search), academic libraries (via paper.search), and general web searches (via web.search) to generate its reports. This creates a surface for indirect prompt injection where malicious instructions embedded in external content could attempt to manipulate the agent's evaluation logic or final report content. * Ingestion points: Step 1 (Patent Search), Step 4 (Paper Search), and Step 5 (Web Search) in SKILL.md. * Boundary markers: The skill uses standard Markdown headers but lacks explicit instructions for the agent to ignore or delimit potentially malicious instructions within the retrieved data. * Capability inventory: The skill has the ability to read from the network and write files to the local environment using the python tool. * Sanitization: The HTML generation script in Step 8.1 of SKILL.md uses html_lib.escape() to sanitize content before writing it to the final report, which mitigates cross-site scripting risks in the output.- [DYNAMIC_EXECUTION]: The skill uses a runtime-generated Python script to process Markdown content into an HTML file. While this is used for legitimate report generation, it involves dynamic interpolation of content into an executable script block. * Evidence: Step 8.1 in SKILL.md defines a Python script template that the agent populates with report data and executes using a python tool to create a localized file at @skill_workspace/.

corp-innovation-brief