The Agent Skills Directory

Indirect Prompt Injection (LOW): The skill possesses a surface for indirect prompt injection by ingesting untrusted data (images and filenames) which are then processed by the system.
Ingestion points: Untrusted data enters via handleImageSelect in references/client.md and the upload.single('image') route in references/server.md.
Boundary markers: The skill lacks explicit instructions to ignore embedded prompts in image metadata; however, it implements robust boundary protections via MIME type validation and random filename generation.
Capability inventory: The skill has capabilities to write to the local filesystem (fs.writeFile), upload to S3 (s3.upload), and execute image processing logic via sharp and squoosh.
Sanitization: Filenames are effectively sanitized by prepending a cryptographically secure random hash in references/storage.md. While the file extension is extracted directly from user input, the randomized prefix prevents directory traversal attacks.

image-compression