The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill instructs users to install the openai-agents Python package. This package is a third-party dependency that does not originate from a known trusted organization or well-known service provider, as seen in the installation commands in SKILL.md and examples/todo-agent.md.
[INDIRECT_PROMPT_INJECTION]: The skill demonstrates tools and architectural patterns that allow agents to ingest and process data from untrusted external sources, which is a prerequisite for indirect prompt injection attacks.
Ingestion points: The fetch_data tool in reference/function-tools.md allows an agent to fetch content from any URL. MCP server integrations shown in templates/agent_mcp.py and examples/todo-agent.md also process data from external endpoints.
Boundary markers: The provided instructions and templates do not include delimiters or specific prompt engineering to warn the model about potentially malicious instructions hidden in the external data.
Capability inventory: The agents shown have access to tools that can perform network requests (httpx), retrieve system state (get_current_time), and perform database-like operations through MCP server tools (e.g., add_task, delete_task).
Sanitization: There is no evidence of sanitization, validation, or filtering of the content retrieved from external sources before it is passed to the agent's context.
[COMMAND_EXECUTION]: The documentation for MCP integration in reference/mcp-integration.md includes the use of MCPServerStdio, which enables an agent to interact with and execute local processes via the command line. While legitimate for local development, this capability must be carefully managed to prevent unauthorized command execution.

openai-agents-sdk