The Agent Skills Directory

[COMMAND_EXECUTION]: The SDK supports arbitrary code execution through its internal tools configuration (internal_tools().code_execution(True)) and the permitted Bash(python *) tool. Reference documentation also includes examples of using the eval() function for tool logic, which can be dangerous if arguments are not strictly validated.
[DATA_EXFILTRATION]: The webhook_tool feature allows agents to send data to external HTTP endpoints. While intended for integrations, it provides a channel for data transfer to third-party services.
[PROMPT_INJECTION]: The skill facilitates building agents that process external data, creating an attack surface for indirect prompt injection.
Ingestion points: Agents ingest data from uploaded files (SKILL.md) and external search results (references/agent-patterns.md).
Boundary markers: Documented examples do not show explicit delimiters or instructions to ignore embedded commands in ingested data.
Capability inventory: Built agents possess significant capabilities including code execution, network requests via webhooks, and potential filesystem access.
Sanitization: The SDK includes a require_approval() mechanism to enable human-in-the-loop confirmation, which is the primary recommended mitigation for sensitive tool executions.

python-sdk