The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill instructs the agent to install a CLI by downloading a script from https://raw.githubusercontent.com/higgsfield-ai/cli/main/install.sh and piping it directly to sh. This pattern executes unverified remote code with the user's shell privileges.
[COMMAND_EXECUTION]: Several bundled Python scripts (brandkit.py, build_brandbook.py, render_brandbook_pdf.py) execute system commands using the subprocess module. These commands include magick, convert, soffice (LibreOffice), rsvg-convert, and pdftoppm. The inputs for these commands are often derived from external assets like SVG or PDF files, which could lead to command injection if paths or filenames are not strictly validated.
[EXTERNAL_DOWNLOADS]: The skill fetches external data from multiple sources, including SVG assets from arbitrary URLs and a PowerPoint template from a hardcoded Google Docs URL. This behavior increases the attack surface for malicious payloads.
[PROMPT_INJECTION]: The skill utilizes an internal 'Design Brain' logic that ingests and analyzes user-supplied brand assets (images, fonts, PDFs, PPTXs). The metadata and content extracted from these files are then used to build prompts for image-generation models (Recraft, Seedream). This pipeline is vulnerable to indirect prompt injection if an attacker embeds malicious instructions within an asset that is subsequently processed and included in a generation prompt.

higgsfield-brandkit