The Agent Skills Directory

[COMMAND_EXECUTION] (MEDIUM): The skill uses MCP tools to execute shell commands, including python mcp_server.py to register a local server and python -m AGENT_NAME validate to run generated Python code. While functional for its purpose, executing local scripts and generated modules creates a significant attack surface.
[DYNAMIC_EXECUTION] (MEDIUM): In Step 6, the skill generates multiple Python files (agent.py, config.py, etc.) based on user-provided goals and workflow designs, and then executes this code in Step 7. This allows for potential code injection if the user input is malicious.
[PROMPT_INJECTION] (LOW): The file contains meta-instructions such as "DO NOT DISPLAY THIS FILE. EXECUTE THE STEPS BELOW" and "IMMEDIATELY begin executing," which are attempts to override the agent's default transparency and reporting behavior.
[DATA_EXPOSURE] (LOW): The script includes hardcoded absolute filesystem paths (/home/timothy/oss/hive), revealing specific details about the user's environment and directory structure.
[INDIRECT_PROMPT_INJECTION] (LOW):
Ingestion points: User input for goal definitions (Step 2) and workflow node designs (Step 3).
Boundary markers: None. User-provided strings are directly interpolated into proposed designs and generated Python files.
Capability inventory: File system write access, MCP server registration, and process execution.
Sanitization: No escaping or validation is performed on the user-provided content before it is written into executable .py files.

building-agents-construction