The Agent Skills Directory

[DYNAMIC_EXECUTION]: The documentation files (references/tool-builder.md and references/agent-patterns.md) provide implementation examples that use eval() to process mathematical expressions in tool handlers and explicitly demonstrate how to enable the codeExecution internal tool capability. While these are presented as SDK features, they introduce potential risks if implemented without strict input validation.
[INDIRECT_PROMPT_INJECTION]: The skill framework facilitates agent interaction with external data sources and tool outputs, creating a vulnerability surface for indirect prompt injection. The documentation does not demonstrate the use of boundary markers or sanitization for these data ingestion points.
Ingestion points: Data enters the agent's context through agent.sendMessage inputs and external tool results provided via agent.submitToolResult (e.g., in references/agent-patterns.md).
Boundary markers: The provided code samples do not include specific delimiters or instructions to ignore instructions embedded within retrieved data.
Capability inventory: The skill framework supports powerful capabilities including local filesystem access (demonstrated via fs in references/agent-patterns.md), network operations (fetch, webhookTool), and arbitrary code execution (internalTools().codeExecution(true)).
Sanitization: The reference patterns do not demonstrate validation or filtering of external content before it is processed by the agent.

javascript-sdk