The Agent Skills Directory

[DYNAMIC_EXECUTION]: The JSXPreview component (documented in references/jsx-preview.md) allows for the dynamic rendering of JSX strings at runtime. This functionality is designed to display AI-generated UI components but creates a surface for arbitrary code execution within the browser if the AI-generated input is not strictly sanitized. The component utilizes the react-jsx-parser library to convert these strings into live React elements.
[INDIRECT_PROMPT_INJECTION]: The skill contains several components, most notably JSXPreview and InlineCitation, that ingest and render data directly from an AI model's output. This creates an indirect prompt injection surface where a subverted model could produce malicious payloads for the UI.
Ingestion points: The jsx prop of the JSXPreview component and the structured source data in InlineCitation.
Boundary markers: None identified; the components do not appear to use explicit delimiters or instructions to ignore commands within the ingested strings.
Capability inventory: These components enable the rendering of complex UI and the execution of logic through the React component tree.
Sanitization: Relies on the default security features of react-jsx-parser and the markdown parser used in MessageResponse.
[DATA_EXFILTRATION]: The SpeechInput component (documented in references/speech-input.md) provides the capability to capture audio directly from the user's microphone. While intended for legitimate voice-to-text features, this represents access to sensitive audio data and local hardware.

ai-elements