The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill fetches files from user-provided URLs using curl and wget. It includes a validation step to ensure only HTTP/HTTPS protocols are used, preventing access to local system files via the file:// protocol.
[COMMAND_EXECUTION]: Shell commands are used for downloading, file management, and Git version control. Filenames are sanitized using sed to restrict characters to alphanumeric, dots, underscores, and dashes, which effectively prevents path traversal and basic command injection via malicious URLs.
[PROMPT_INJECTION]: The skill analyzes the text content of downloaded PDFs and images to generate descriptive filenames. This creates a surface for indirect prompt injection, where a malicious document could contain instructions designed to influence the agent's behavior during the analysis phase.
Ingestion points: Files downloaded from external URLs to the 05_Attachments/ directory.
Boundary markers: None explicitly provided to isolate the analyzed content from the agent's instructions.
Capability inventory: wget, curl, ls, git, pdftotext, and vision-based analysis tools.
Sanitization: Filenames are sanitized, but the internal content of files is processed as untrusted input for natural language analysis.

download-attachment