The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill provides templates and patterns for building AI agents that ingest untrusted user data via agent.sendMessage() (as seen in SKILL.md and references/react-integration.md). These agents possess high-privilege capabilities including network operations via webhookTool, file system access through SDK file handling, and code execution via internalTools. The provided examples do not include boundary markers or input sanitization, exposing a surface for indirect prompt injection from processed data.
Ingestion points: agent.sendMessage(input) calls in SKILL.md and React hook examples.
Boundary markers: Absent in standard agent initialization and messaging examples.
Capability inventory: The SDK supports webhookTool (network), uploadFile/readFileSync (file access), and codeExecution(true) (internal tool).
Sanitization: No sanitization or escaping of input data is demonstrated in the integration examples.
[DYNAMIC_EXECUTION]: The documentation in references/tool-builder.md includes a code snippet showing the use of eval() to process expressions within a tool handler (result = eval(call.args.expression)). While provided as a developer reference for client-side tool execution, it promotes an insecure coding practice where untrusted input is directly executed.
[EXTERNAL_DOWNLOADS]: The skill instructs the user to download and install additional components via the CLI command npx skills add belt-sh/cli.

javascript-sdk