The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill establishes several vulnerability surfaces where untrusted data (media files, documents) is ingested into the application context. If these files contain malicious instructions or adversarial content, they could influence the agent's behavior during processing.
Ingestion points: Multiple file upload patterns (Basic, Advanced, Image-specific) in SKILL.md and implementation guides in references/upload-patterns.md.
Boundary markers: The provided examples lack explicit sanitization or instructions to ignore embedded content within processed documents.
Capability inventory: The skill includes capabilities for client-side editing (react-easy-crop), metadata extraction, and document conversion.
Sanitization: Significant risk identified in references/office-viewer.md which demonstrates converting Word documents to HTML using mammoth.js and rendering them via dangerouslySetInnerHTML without visible sanitization logic.
[DYNAMIC_EXECUTION]: The skill uses dynamic execution and rendering techniques for media content.
Evidence: references/office-viewer.md exhibits the use of dangerouslySetInnerHTML to render dynamically generated HTML from Word document buffers.
Evidence: examples/pdf-react.tsx loads an external worker script for PDF rendering from a CDN (cdnjs.cloudflare.com), representing dynamic loading from an external source.
[EXTERNAL_DOWNLOADS]: The skill references and integrates multiple external libraries and cloud services.
Evidence: Recommends installation of various npm packages including video.js, react-pdf, wavesurfer.js, and mammoth.js.
Evidence: Provides integration patterns for well-known services such as AWS S3, Cloudflare R2, Cloudinary, and Mux.
Note: All identified external references target well-known services or trusted organizations, which are considered safe per standard auditing rules.

managing-media