The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The scripts/package-loader.mjs script implements a bootstrap mechanism that automatically installs missing Node.js dependencies from the npm registry into a temporary directory and executes them. While it uses security measures like --ignore-scripts to prevent the execution of malicious installation hooks and requires user confirmation, it fundamentally allows the skill to download and run external code at runtime.
[COMMAND_EXECUTION]: The scripts/extract-audio-data.py script executes the ffmpeg command-line utility via subprocess.run() to process audio files. It takes user-provided file paths as arguments, which could be a vector for command manipulation if not handled strictly by the underlying OS.
[DYNAMIC_EXECUTION]: The skill uses dynamic import() calls within scripts/package-loader.mjs to load JavaScript modules from computed paths in a temporary node_modules directory created during the bootstrap process. Additionally, scripts/contrast-report.mjs uses browser automation to execute code within a page context for accessibility auditing.
[EXTERNAL_DOWNLOADS]: The skill is designed to download various Node.js packages (e.g., @hyperframes/producer, @hyperframes/core, sharp) from the npm registry if they are not already present in the environment.
[INDIRECT_PROMPT_INJECTION]: The design picker workflow (references/design-picker.md) generates an HTML file (.hyperframes/pick-design.html) that interpolates model-generated content and user-provided text (like headlines). This content is rendered in a local browser session via innerHTML. Although the instructions warn the model against generating scripts or event handlers, this creates a potential cross-site scripting (XSS) surface if a malicious prompt influences the generated HTML.

hyperframes-creative