The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill instructions direct the user to perform a git clone of an external repository (github.com/img2threejs/img2threejs.git) into the user's local directory and then execute Python scripts from this repository using uv run. Since this repository is not affiliated with a trusted organization, this pattern allows for the execution of unverified third-party code.
[DYNAMIC_EXECUTION]: The script scripts/render_model.py dynamically generates TypeScript and HTML files (main.ts and index.html) at runtime and writes them to a hidden directory. These files are then executed within a Vite and Playwright environment to render 3D assets, creating a vector for hidden code execution.
[DYNAMIC_EXECUTION]: The script scripts/align_framing.py modifies the Python search path (sys.path) at runtime to import modules from the external, cloned repository. This dynamic loading of modules from computed paths allows for the execution of arbitrary code from the external source.
[COMMAND_EXECUTION]: The scripts/render_model.py script uses subprocess.Popen to launch shell commands, specifically npx vite. While used here for a development server, spawning shell processes to execute commands derived from the skill environment presents a risk of command injection or unauthorized system access.
[INDIRECT_PROMPT_INJECTION]: The skill processes external reference images and requires the agent to manually author and fill complex JSON specifications (spec.json, assessment.json). This workflow ingests untrusted data into the agent's context without explicit boundary markers or sanitization, creating an attack surface where instructions embedded in images or crafted data could influence the agent's code generation steps.

image-to-threejs