The Agent Skills Directory

[DYNAMIC_EXECUTION]: The documentation provides code examples that use the Python eval() function to process model-generated arguments. In references/tool-builder.md, a calculator tool is implemented by calling eval(call.args['expression']), which poses a risk of arbitrary code execution if the input is influenced by malicious instructions.- [DYNAMIC_EXECUTION]: The SDK supports enabling built-in code execution capabilities for agents via internal_tools().code_execution(True), as seen in SKILL.md and references/agent-patterns.md. This allows the agent to generate and run code dynamically during execution.- [INDIRECT_PROMPT_INJECTION]: The skill facilitates agents that ingest untrusted data from multiple sources, creating a surface for indirect prompt injection.
Ingestion points: Data enters the system via agent.send_message(), file uploads, and external inputs described in SKILL.md and references/files.md.
Boundary markers: The SDK supports human-in-the-loop validation using require_approval(), though developers must implement these checks manually.
Capability inventory: Agents can perform network requests (via webhook_tool), execute code, and search the web, providing high-impact targets for injection attacks.
Sanitization: The documentation emphasizes approval workflows for sensitive tool calls.- [EXTERNAL_DOWNLOADS]: The skill instructs users to download and install external dependencies, including the inferencesh Python package and the belt-sh/cli tool via npx.- [COMMAND_EXECUTION]: The skill's documentation and configuration involve executing shell commands such as pip install and python to manage dependencies and run applications.

python-sdk