The Agent Skills Directory

[DYNAMIC_EXECUTION]: Documentation examples in references/agent-patterns.md and references/tool-builder.md demonstrate the use of the Python eval() function to process input provided by the AI agent (e.g., within a 'calculate' tool). Executing eval() on strings generated or influenced by the agent allows for arbitrary code execution if the input is not strictly validated or sanitized.
[INDIRECT_PROMPT_INJECTION]: The skill exposes a significant attack surface for indirect prompt injection due to its core functionality of processing external data with high-privilege tools.
Ingestion points: The SDK ingests untrusted data via file uploads (references/files.md), user-provided messages, and outputs from external tools like web search assistants or webhooks.
Boundary markers: The provided examples do not demonstrate the use of delimiters or specific instructions to the model to ignore instructions embedded within external data.
Capability inventory: The skill enables agents with powerful capabilities, including arbitrary code execution (internal_tools().code_execution(True)), file system operations (e.g., the delete_file example in references/agent-patterns.md), and the ability to trigger external HTTP requests via webhook_tool.
Sanitization: Examples involving the execution of agent-provided logic (such as the eval() pattern) lack input validation or filtering of external content.
[COMMAND_EXECUTION]: The skill configuration in SKILL.md explicitly allows the agent to execute Python commands and scripts using the Bash(python *) tool, which is a high-privilege capability necessary for the SDK's intended purpose but increases the impact of successful injections.
[EXTERNAL_DOWNLOADS]: The skill facilitates the installation and use of the inferencesh Python package from standard registries, which is the official SDK for the service provided by the author.

python-sdk