The Agent Skills Directory

Command Execution: The skill facilitates the management and execution of commands within isolated microVMs via the sbx CLI. This includes starting agents, running ad-hoc commands with sbx exec, and copying files between the host and sandbox environment.
Indirect Prompt Injection: A potential attack surface exists where an agent could ingest untrusted data through mounted host directories as described in SKILL.md. The skill documentation mitigates this risk by providing explicit warnings against mounting secrets or credentials. The capability inventory includes lifecycle management and command execution tools as defined in skill.yaml, while no specific sanitization or boundary markers are defined for the mounted data itself.
Privileged Access: The analysis notes that sbx exec supports the --privileged flag, which can grant extended capabilities to processes within the microVM. This is documented as a standard feature for administrative tasks in isolated environments.
Secure Resource Management: The skill emphasizes safe practices for resource teardown, such as using sbx prune with age filters and dry-runs, and ensuring Git commits are fetched before removing clone-mode sandboxes to prevent data loss.

docker-sandboxes-lifecycle