The Agent Skills Directory

[COMMAND_EXECUTION]: Documentation examples in references/agent-patterns.md and references/tool-builder.md demonstrate the use of Python's eval() function to process user-supplied expressions for a 'calculator' tool. Using eval() on untrusted input is a significant security risk that can lead to arbitrary code execution within the environment where the SDK is running.
[EXTERNAL_DOWNLOADS]: The skill instructs users to install external dependencies such as belt-sh/cli via npx skills add belt-sh/cli and the inferencesh Python package. While these are presented as official project components, users should verify the integrity of these packages before installation.
[DATA_EXFILTRATION]: The SDK facilitates the upload of local files and environment variables (like INFERENCE_API_KEY) to the inference.sh platform. This is the intended primary purpose of the skill, but users should ensure they do not accidentally upload sensitive local files through the automatic upload features described in references/files.md.
[DYNAMIC_EXECUTION]: The skill documentation highlights a 'Code Execution Pattern' and 'Internal Tools' configuration (.code_execution(True)) that allows agents to write and run code. This provides a powerful capability that requires strict oversight, particularly when the agent processes untrusted data.
[INDIRECT_PROMPT_INJECTION]: The skill describes building agents that ingest data from external sources like web searches and app outputs (references/agent-patterns.md). This creates a surface for indirect prompt injection where malicious content in a search result or app response could influence the agent's behavior. The documentation correctly mitigates this risk by recommending human-in-the-loop approval workflows for sensitive actions.

python-sdk