The Agent Skills Directory

[COMMAND_EXECUTION]: The skill provides instructions for the agent to use the docker CLI and the docker sandbox plugin to create, execute, and manage isolated virtual machine environments on the host system.
[EXTERNAL_DOWNLOADS]: The documentation includes examples of using curl, wget, and npm install within the sandbox to fetch external resources, which aligns with the skill's purpose of safely exploring untrusted code.
[INDIRECT_PROMPT_INJECTION]: The skill's primary workflow involves mounting host directories into the sandbox via virtiofs. While this creates a surface for the agent to ingest potentially malicious data from the workspace, the sandbox itself is the recommended mitigation for isolating such risks.
[SAFE]: The skill follows security best practices by encouraging environment isolation and providing detailed instructions for network proxying and allowlist-based traffic control for AI workloads.

docker-sandbox