The Agent Skills Directory

[DYNAMIC_EXECUTION]: The documentation includes an example in references/tool-builder.md that uses eval() to process expressions generated by an AI agent, which is a significant security risk that could lead to arbitrary code execution if the agent is manipulated.
[DYNAMIC_EXECUTION]: The skill facilitates the creation of agents with a "Code Execution" capability via internalTools().codeExecution(true), allowing agents to write and run code in the execution environment.
[INDIRECT_PROMPT_INJECTION]: The skill creates a surface for indirect prompt injection by demonstrating patterns where agents ingest data from untrusted sources, such as web search results or uploaded files, and process them without explicit boundary markers or sanitization logic.
Ingestion points: Data enters the agent's context through user messages (agent.sendMessage), web search tool results (tavily/search-assistant), and user-uploaded file content.
Boundary markers: No delimiters or "ignore instructions" warnings are shown in the provided SDK implementation examples.
Capability inventory: The skill has access to Bash for package management and script execution, file system read/write via Node.js, and outbound network communication via webhooks.
Sanitization: No input validation or escaping of external content is demonstrated in the provided examples.
[EXTERNAL_DOWNLOADS]: The skill documentation prompts the installation of the @inferencesh/sdk package and suggests adding the external belt-sh/cli skill, introducing dependencies from non-trusted repositories.

javascript-sdk