The Agent Skills Directory

[PROMPT_INJECTION]: The documentation describes patterns for building AI agents that process untrusted data from external sources like emails and webhooks. Multiple examples demonstrate direct interpolation of this content into AI models without sufficient sanitization or the use of boundary delimiters, creating a potential surface for indirect prompt injection.
Ingestion points: references/agents-sdk/patterns.md (onEmail handler) and references/email-workers/api.md (ForwardableEmailMessage).
Boundary markers: Absent in the provided implementation examples.
Capability inventory: The examples show agents with access to persistent storage (SQL), task scheduling, and network operations.
Sanitization: Not explicitly shown in the AI integration code snippets.
[EXTERNAL_DOWNLOADS]: The documentation includes examples that fetch configuration or setup scripts from external domains.
Evidence: references/sandbox/patterns.md contains a command to install an IDE using curl -fsSL https://code-server.dev/install.sh | sh. While the source is a well-known service, it follows the pattern of remote script execution.
[REMOTE_CODE_EXECUTION]: The skill documents the Cloudflare Sandbox product, which is designed for dynamic code execution in isolated containers. The examples provided teach how to execute arbitrary shell commands and Python code within these environments.
Evidence: references/sandbox/api.md documents sandbox.exec() and ctx.runCode() methods for executing dynamic logic.

cloudflare-deploy