The Agent Skills Directory

[COMMAND_EXECUTION]: The documentation describes enabling internal code execution capabilities (code_execution(True)) for AI agents. This allows the agent to generate and execute its own code to complete tasks, which poses a security risk if the agent processes untrusted inputs.
[REMOTE_CODE_EXECUTION]: Documentation examples in references/tool-builder.md demonstrate the use of the eval() function to handle mathematical expressions within tool handlers. This is an unsafe programming practice that allows for arbitrary code execution if the agent is manipulated into generating a malicious payload.
[EXTERNAL_DOWNLOADS]: The skill requires the installation of the inferencesh Python package and references other libraries such as requests, aiohttp, and fastapi for various integration and processing patterns.
[PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection (Category 8).
Ingestion points: Agents can ingest untrusted data through local file uploads (client.upload_file), file attachments in messages, and results from external search tools.
Boundary markers: There are no instructions for using delimiters or warnings to ignore embedded commands when the agent processes external data.
Capability inventory: The SDK provides powerful tools including internal code execution, network access via webhooks, and file system management.
Sanitization: No sanitization or validation techniques are demonstrated before processing untrusted content or executing tool calls based on agent output.

python-sdk