The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill documentation in runtime/reference-assets/upstream-js/runjs/context/import-async.md and require-async.md provides instructions and examples for using ctx.importAsync() and ctx.requireAsync() to dynamically load and execute code from remote CDNs such as esm.sh and jsdelivr.net. This allows the agent or generated scripts to run arbitrary remote code.
[DYNAMIC_EXECUTION]: Hard Rule 16 in SKILL.md and the authoring loop in references/js.md explicitly direct the agent to generate JavaScript source code and write it into code or script fields within the NocoBase UI configuration. This code is then executed in a browser-based sandbox environment.
[DATA_EXFILTRATION]: Documentation in runtime/reference-assets/upstream-js/runjs/context/get-var.md confirms that generated scripts have access to ctx.token (the user's API Token). When combined with the ctx.request() capability (documented in request.md), this provides a mechanism for scripts to harvest and transmit authentication credentials to external endpoints.
[INDIRECT_PROMPT_INJECTION]: Hard Rule 4 in SKILL.md instructs the agent to "treat collection field description as active form-authoring input" and use semantic extraction for behavior rules. This creates an attack surface where a malicious user could embed prompt injection instructions in database metadata (field descriptions) to manipulate the agent's UI construction or code generation logic.

nocobase-ui-builder