The Agent Skills Directory

[PROMPT_INJECTION]: The skill architecture documents the ingestion of untrusted external data via ScrapeWebsiteTool and SerperDevTool as shown in SKILL.md. This represents an indirect prompt injection surface where malicious content from external websites could influence agent behavior. The guide does not provide examples of boundary markers or sanitization logic to isolate this untrusted content. Agents described in references/advanced_patterns.md possess significant capabilities, including file system access and command execution via MCPTool, increasing the potential impact of successful injection.
[REMOTE_CODE_EXECUTION]: A persistence pattern described in references/advanced_patterns.md demonstrates the use of the pickle module for saving and loading workflow states. The documentation shows pickle.load() being used to restore state from workflow_state.pkl, which is an unsafe operation if the file originates from or can be modified by an untrusted source, as it allows for arbitrary code execution.
[COMMAND_EXECUTION]: The skill provides integration examples for MCPTool in references/advanced_patterns.md, which enables agents to execute external shell commands (e.g., using npx to launch MCP servers). This functionality provides agents with a mechanism to run arbitrary code or binaries within the execution environment.

crewai-developer