The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill facilitates the installation of the @cloudflare/sandbox package via NPM and utilizes official Docker images from Docker Hub (docker.io/cloudflare/sandbox). It also retrieves documentation and code examples from Cloudflare's official domains and GitHub repositories. All external resources are from well-known services or trusted organizations.
[COMMAND_EXECUTION]: The skill provides methods such as sandbox.exec() to run shell commands within the isolated container environment. This is the intended functionality of the sandbox SDK for building CI/CD systems or interactive dev environments.
[DYNAMIC_EXECUTION]: The skill includes the sandbox.runCode() method, which allows for the runtime interpretation of Python, JavaScript, and TypeScript. This is designed for executing LLM-generated code within a secure, isolated context.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to process and execute untrusted or LLM-generated code, creating a potential surface for indirect prompt injection.
Ingestion points: Untrusted code or command strings are passed to the runCode and exec methods as shown in SKILL.md and references/api-quick-ref.md.
Boundary markers: The SDK itself serves as the isolation boundary, keeping execution limited to the containerized environment.
Capability inventory: The skill allows for file system operations (readFile, writeFile), network port exposure (exposePort), and arbitrary command execution (exec) within the sandbox.
Sanitization: The skill relies on the underlying Cloudflare Workers infrastructure to provide secure isolation for the executed code.

sandbox-sdk