The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructions in SKILL.md direct the agent to use sudo apt install on Linux systems to install Python and venv components during the bootstrap process.- [EXTERNAL_DOWNLOADS]: The skill downloads approximately 3GB of machine learning model weights and supporting files from Microsoft's repository on Hugging Face (microsoft/OmniParser-v2.0). It also installs numerous Node.js and Python dependencies from official registries.- [REMOTE_CODE_EXECUTION]: The skill executes third-party code through the installation and runtime usage of several large libraries including Playwright, Transformers, and Torch. It also runs a Python-based machine learning handler to process user-provided images.- [PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection. It extracts text from user-supplied design images via OCR (OmniParser). This extracted text is then returned to the agent's context to guide UI modifications, creating a vector where malicious text in a design file could influence agent actions (Category 8).

frontend-audit