The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill documentation includes instructions for installing external software, specifically the @inferencesh/sdk npm package and the belt-sh/cli skill, which are used to interact with the platform services.\n- [COMMAND_EXECUTION]: The skill is granted permissions via allowed-tools to run npm, npx, node, pnpm, and yarn. This enables the agent to perform local environment setup and execute JavaScript code, which could be misused if the agent is compromised by external data.\n- [REMOTE_CODE_EXECUTION]: Documentation examples in references/tool-builder.md and references/agent-patterns.md demonstrate processing tool arguments using eval(). While presented as examples for a calculator tool, this pattern promotes unsafe execution of arbitrary strings which can lead to code injection vulnerabilities if implemented by users without strict sanitization.\n- [PROMPT_INJECTION]: The skill architecture supports building agents that process data from external, untrusted sources (e.g., via web search, webhooks, or third-party apps), creating a vulnerability surface for indirect prompt injection.\n
Ingestion points: External data enters through tools like webhookTool or internalTools().webSearch() as shown in references/tool-builder.md.\n
Boundary markers: There are no examples of using boundary markers or specific negative constraints to prevent the agent from following instructions embedded within processed data.\n
Capability inventory: The skill allows high-privilege operations including shell access (via frontmatter), file system writes (references/files.md), and outbound network requests.\n
Sanitization: The provided code snippets do not include logic for sanitizing or escaping content retrieved from external sources before it is interpreted by the agent.

javascript-sdk