The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The script scripts/package-loader.mjs implements a bootstrapping mechanism that can dynamically install Node.js packages from the NPM registry using npm install if they are not resolved locally. While it includes mitigations such as disabling lifecycle scripts via --ignore-scripts, requiring pinned versions, and a TTY-check for user confirmation, the ability to download and execute code at runtime from external registries is a significant risk vector.
[DYNAMIC_EXECUTION]: In scripts/package-loader.mjs, the script uses the dynamic import() statement on computed paths resolved at runtime. This is used to load dependencies that may have been just installed by the bootstrapper.
[COMMAND_EXECUTION]: The Python script scripts/extract-audio-data.py uses subprocess.run to execute the ffmpeg command-line utility for audio decoding. This script accepts a user-provided file path as input, which is passed to the command line.
[INDIRECT_PROMPT_INJECTION]: The templates/design-picker.html file is used as a template for an interactive tool where architectural previews are injected via innerHTML. This creates a vulnerability surface for indirect prompt injection if the metadata (__ARCHITECTURES_JSON__) processed by the agent contains malicious HTML or scripts. Documentation in references/design-picker.md acknowledges this risk and provides guidelines to avoid script tags, but the surface remains present.

hyperframes-creative