The Agent Skills Directory

[COMMAND_EXECUTION]: The skill provides instructions for system-level configuration, including using sudo to install packages and restart the Docker daemon in references/01-gpu-setup.md. It also documents the docker agent feature in SKILL.md, which utilizes an execute_command tool for executing tasks within isolated sandbox environments.
[EXTERNAL_DOWNLOADS]: The references/01-gpu-setup.md file contains instructions to download repository keys and configuration files from NVIDIA's official GitHub-hosted service (nvidia.github.io). This is a standard and necessary step for enabling hardware acceleration and originates from a well-known technology provider.
[PROMPT_INJECTION]: The Docker Agent configuration described in SKILL.md represents a surface for indirect prompt injection as it facilitates the processing of user-provided goals using high-capability tools.
Ingestion points: Untrusted instructions enter the agent context via the command-line interface as illustrated in SKILL.md (e.g., docker agent run -f agent.yaml "Fix the bug...").
Boundary markers: The example agent.yaml provided in SKILL.md does not specify delimiters or security instructions to prevent the agent from obeying commands embedded in files it may read.
Capability inventory: According to SKILL.md, the agent is configured with powerful capabilities including execute_command and write_file.
Sanitization: No input sanitization, filtering, or validation logic is documented within the skill's instructions or example configurations.

docker-ai-ml