The Agent Skills Directory

[DYNAMIC_EXECUTION]: The reference files (references/agent-patterns.md and references/tool-builder.md) contain code examples demonstrating the use of eval() to process expressions generated by AI models. While provided as a coding pattern for a calculator tool, developers should be aware that using eval on untrusted model output is a high-risk security practice.
[INDIRECT_PROMPT_INJECTION]: The SDK facilitates the creation of agents that ingest data from external sources such as web searches, file uploads, and webhook responses, establishing an attack surface for indirect prompt injection.
Ingestion points: Ingestion occurs through file handling (uploadFile), web search capabilities (webSearch), and webhook integrations (webhookTool) defined in the documentation.
Boundary markers: The provided examples do not explicitly demonstrate the use of strict delimiters to isolate untrusted data from the core system prompt.
Capability inventory: The SDK exposes powerful capabilities across its scripts including file system access, network requests via webhooks, and cloud-side code execution (internalTools().codeExecution(true)).
Sanitization: The skill documents and encourages the use of human-in-the-loop approval workflows (requireApproval()) and server-side proxies to mitigate the risk of unauthorized or malicious operations.

javascript-sdk