The Agent Skills Directory

[COMMAND_EXECUTION]: Usage of the eval() function in CalculatorTool examples found in SKILL.md and references/tools.md. The implementation in SKILL.md lacks any input validation, while the version in references/tools.md uses a superficial isalpha() check that is insufficient to prevent all forms of malicious code execution.
[COMMAND_EXECUTION]: Documentation includes a CodeInterpreterTool for executing Python code. While described as sandboxed, it provides a high-privilege capability for agents to run arbitrary scripts.
[DATA_EXFILTRATION]: The skill documents several tools for file system and database interaction, including FileReadTool, FileWriterTool, DirectoryReadTool, MySQLSearchTool, and PostgreSQLTool. These tools allow agents to read and modify local files or query remote databases.
[EXTERNAL_DOWNLOADS]: Extensive support for web scraping and remote data retrieval via tools like ScrapeWebsiteTool, SeleniumScrapingTool, and FirecrawlScrapeWebsiteTool. These tools enable the agent to fetch and process content from arbitrary external URLs.
[INDIRECT_PROMPT_INJECTION]: The skill possesses a significant attack surface for indirect prompt injection due to its ability to ingest data from untrusted sources (web, PDF, CSV, etc.).
Ingestion points: ScrapeWebsiteTool, PDFSearchTool, FileReadTool, YoutubeVideoSearchTool (documented in references/tools.md).
Boundary markers: Absent; examples show direct interpolation of variables into task descriptions without delimiters.
Capability inventory: CodeInterpreterTool, FileWriterTool, MySQLSearchTool (documented in references/tools.md).
Sanitization: No mention of input sanitization or filtering for ingested content before it is processed by agents.

crewai-multi-agent