The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill demonstrates patterns for ingesting and displaying untrusted data that could be exploited for malicious input.
Ingestion points: event.target.files in references/2024-09-17-stimulus-image-upload-previews.md (file names) and region.content.innerText in references/2024-07-02-stimulus-wavesurfer-add-markers.md (audio marker text).
Boundary markers: Absent. Data is interpolated directly into templates without explicit separators or instructions to ignore embedded content.
Capability inventory: The skill snippets perform DOM manipulation via this.previewsTarget.appendChild in references/2024-09-17-stimulus-image-upload-previews.md and this.listTarget.insertAdjacentHTML in references/2024-07-02-stimulus-wavesurfer-add-markers.md.
Sanitization: Absent. The implementation relies on simple string replacement or external library behavior without explicit escaping or sanitization of external input.
[DYNAMIC_EXECUTION]: Multiple reference files demonstrate runtime HTML construction using manual string manipulation, which is an unsafe practice when handling external data.
In references/2024-09-17-stimulus-image-upload-previews.md, the appendPreviews method replaces placeholders in a template with file.name and assigns the result to innerHTML. This pattern is vulnerable to XSS if a user selects a file with a maliciously crafted name.
In references/2024-07-02-stimulus-wavesurfer-add-markers.md and references/2024-07-30-stimulus-wavesurfer-remove-markers.md, the code uses .replace() to inject library-provided content into HTML strings before calling insertAdjacentHTML. Without proper escaping, this allows for the injection of arbitrary HTML/JS if the input content is controlled by an attacker.

hwc-media-content