The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: Fetches documentation, API references, and examples from Cloudflare's official developer portal and GitHub repositories.
[COMMAND_EXECUTION]: Enables execution of shell commands within the isolated sandbox environment using the sandbox.exec interface.
[DYNAMIC_EXECUTION]: Facilitates runtime code execution for Python, JavaScript, and TypeScript via the sandbox.runCode and createCodeContext methods.
[INDIRECT_PROMPT_INJECTION]: The skill identifies the risk of processing untrusted LLM-generated input and provides comprehensive mitigations:
Ingestion points: Untrusted code or commands are passed to the runCode or exec methods.
Boundary markers: Recommends the use of inputSchema to validate data before processing.
Capability inventory: Restricts execution to isolated containers with configurable resource limits (CPU, memory, timeout).
Sanitization: Enforces least-privilege network policies by routing all external requests through an allowed host callback pattern.

sandbox-sdk