The Agent Skills Directory

[COMMAND_EXECUTION] (HIGH): The documentation in references/tool-builder.md explicitly demonstrates using Python's eval() function to handle tool arguments (e.g., eval(call.args['expression'])). Since these arguments are generated by the LLM based on user prompts, this allows for arbitrary command execution.\n- [REMOTE_CODE_EXECUTION] (HIGH): The internal_tools().code_execution(True) pattern shown in references/agent-patterns.md allows agents to write and run code. In a production environment, this capability requires strict sandboxing which is not discussed in the reference.\n- [DATA_EXFILTRATION] (MEDIUM): The webhook_tool capability in references/tool-builder.md enables agents to transmit data to arbitrary external URLs. This can be abused to leak sensitive information if an agent is tricked into sending secrets as webhook parameters.\n- [PROMPT_INJECTION] (LOW): The RAG and Webhook patterns create a surface for indirect prompt injection where untrusted external data influences agent behavior.\n
Ingestion points: Web search results (agent-patterns.md) and Webhook responses (tool-builder.md).\n
Boundary markers: Absent in provided system prompts.\n
Capability inventory: Filesystem deletion (delete_file), arbitrary code execution (eval), and network egress (webhook_tool).\n
Sanitization: None demonstrated; external data is processed directly by the LLM.

python-sdk