The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructions utilize docker run, docker build, and docker exec to manage containerized execution. These commands are used to install dependencies and execute research scripts in a controlled environment.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and run untrusted code from external research repositories, which constitutes a surface for indirect prompt injection. The primary security measure provided is the use of Docker for environment isolation.
Ingestion points: Third-party research repositories, project files, and requirements.txt manifests.
Boundary markers: Relies on Docker container boundaries; documentation suggests using --network none for additional isolation.
Capability inventory: Shell command execution and script invocation within the container.
Sanitization: No explicit sanitization of the external code is described, as the skill focuses on sandboxing the execution.
[EXTERNAL_DOWNLOADS]: The skill fetches container images from well-known and official sources, including Docker Hub official images for Python, Node.js, Ubuntu, R, and Julia, as well as verified organization images from PyTorch and TensorFlow.

docker