The Agent Skills Directory

[DYNAMIC_EXECUTION]: The reference guides (specifically references/tool-builder.md and references/agent-patterns.md) provide code examples that use the JavaScript eval() function to implement a calculator tool. While presented as an example, this pattern facilitates arbitrary code execution if implemented as shown without rigorous input sanitization.
[INDIRECT_PROMPT_INJECTION]: The SDK facilitates the creation of agents that ingest and process untrusted data from users, web search results, and file uploads, creating a vulnerability surface for indirect prompt injection.
Ingestion points: Untrusted data enters the agent context through agent.sendMessage inputs, appTool outputs, webSearch results, and file content processing described in SKILL.md and references/files.md.
Boundary markers: Documentation examples do not demonstrate the use of delimiters or isolation prompts to separate untrusted data from instructions.
Capability inventory: Agents built using this SDK are demonstrated with access to sensitive capabilities including shell execution (Bash tools), file system operations (fs), and network communication.
Sanitization: Provided examples do not include evidence of input validation or sanitization for data interpolated into agent prompts or passed to tool handlers.
[EXTERNAL_DOWNLOADS]: The documentation recommends the installation of an external utility skill (belt-sh/cli) via the platform's extension mechanism. Users should exercise caution when adding third-party skills that extend platform capabilities.

javascript-sdk