The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The documentation in references/tool-builder.md repeatedly provides code examples (under 'Basic Handler' and 'Complete Example') that use the unsafe eval() function to process expression inputs from the AI agent. This pattern guides developers to create a Remote Code Execution (RCE) vulnerability where an attacker could execute arbitrary code on the host machine by manipulating the agent's output via prompt injection.
[EXTERNAL_DOWNLOADS]: The skill facilitates the installation and use of the @inferencesh/sdk package from the npm registry. It also recommends adding the belt-sh/cli skill using npx in the SKILL.md file.
[PROMPT_INJECTION]: The skill defines an architecture for agents that process untrusted data, creating an attack surface for indirect prompt injection.
Ingestion points: Data enters the agent context through agent.sendMessage() and outputs from tools like appTool or webhookTool as seen in references/agent-patterns.md and references/tool-builder.md.
Boundary markers: The provided code examples do not demonstrate the use of delimiters (e.g., XML tags or triple quotes) or 'ignore' instructions to isolate untrusted data during interpolation.
Capability inventory: Agents built with this SDK can execute local code (if enabled), access the file system via uploadFile, and perform network operations through the SDK or webhooks.
Sanitization: There is no evidence of input validation or output sanitization in the integration patterns provided.

javascript-sdk