The Agent Skills Directory

[COMMAND_EXECUTION]: The skill provides Python code snippets demonstrating a minimal agent loop that uses the eval() function to implement a calculator tool. While the code attempts to restrict the execution environment using {"__builtins__": {}}, this pattern is noted as an educational example and the skill explicitly provides safety checklists and human-in-the-loop patterns to mitigate risks in production.
[EXTERNAL_DOWNLOADS]: The skill references and provides commands to clone a repository from GitHub (datawhalechina/Agent-Learning-Hub) and links to official documentation from trusted sources like Anthropic, OpenAI, and the Model Context Protocol website. These resources are standard for the skill's educational purpose.
[DATA_EXFILTRATION]: No patterns of sensitive data collection or unauthorized network transmission were found. Code examples use environment variables for API keys, which is a recommended security practice.
[INDIRECT_PROMPT_INJECTION]: The skill outlines architectures for agents that ingest untrusted external data (via RAG and browser automation), creating a potential attack surface.
Ingestion points: Data loading via SimpleDirectoryReader and web scraping via browser-use.
Boundary markers: Recommends citation verification and grounding mechanisms to prevent hallucinations or injections.
Capability inventory: Includes capabilities for file reading, command execution, and browser interaction.
Sanitization: Provides specific examples for input validation using Pydantic models to ensure tool arguments are well-formed.

datawhale-agent-learning-hub