The Agent Skills Directory

[COMMAND_EXECUTION]: The instructions explicitly inform the agent that it can use sudo to execute commands as root within the sandbox environment.\n- [COMMAND_EXECUTION]: The skill utilizes the agentbox-ctl checkpoint command to persist the state of the sandbox's writable layer, which captures all filesystem changes made by the agent.\n- [DATA_EXFILTRATION]: The skill describes mechanisms (carry: and agentbox-ctl cp) to move files from the host machine into the sandbox. Examples provided include potentially sensitive data like ~/.claude (which may contain authentication tokens) and files from ~/secrets/.\n- [PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection as it instructs the agent to parse and act upon untrusted project files.\n- [PROMPT_INJECTION]: Ingestion points: Manifest files in /workspace such as package.json, pyproject.toml, Makefile, and docker-compose.yaml.\n- [PROMPT_INJECTION]: Boundary markers: None provided; the agent is not instructed to use delimiters or ignore instructions found in these files.\n- [PROMPT_INJECTION]: Capability inventory: Root access via sudo, writing configuration files, and state persistence via checkpoint.\n- [PROMPT_INJECTION]: Sanitization: None; the agent uses extracted data directly to generate the system configuration.

agentbox-setup