The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill provides patterns for agents to ingest untrusted data, which can lead to indirect prompt injection attacks where malicious instructions are hidden in external data.
Ingestion points: Data enters through onChatMessage (references/streaming-chat.md), onEmail (references/email.md), and onRequest (references/webhooks-push.md).
Boundary markers: Absent; code snippets show direct interpolation of messages without specialized delimiters or "ignore" instructions for embedded content.
Capability inventory: The SDK provides powerful tools across multiple files, including this.sql (references/state-scheduling.md), createBrowserTools (references/browse-the-web.md), and this.mcp.getAITools() (references/mcp.md).
Sanitization: Absent; no examples of escaping or validation of untrusted input before processing are provided in the references.
[DYNAMIC_EXECUTION]: The skill describes "Codemode" and "Browser tools" which allow LLMs to generate and execute JavaScript code at runtime within isolated environments.
Evidence: references/codemode.md details the DynamicWorkerExecutor which runs generated JavaScript in a sandbox. references/browse-the-web.md describes LLMs writing async JS IIFEs for browser interaction.
[EXTERNAL_DOWNLOADS]: The documentation provides instructions to install several npm packages from the official registry.
Packages: agents, @cloudflare/ai-chat, ai, @ai-sdk/react, @cloudflare/codemode, zod, @modelcontextprotocol/sdk, @cloudflare/workers-oauth-provider, web-push, @cloudflare/think, @cloudflare/voice, postal-mime.
Note: The majority of packages belong to a well-known service provider or are established community libraries.
[COMMAND_EXECUTION]: The SKILL.md file includes commands for environment verification and package installation.
Evidence: npm ls agents and npm install agents in the verification section.

agents-sdk