The Agent Skills Directory

[COMMAND_EXECUTION]: Spawns the host-level grok CLI to perform shell commands and file edits in any specified host directory.- [COMMAND_EXECUTION]: Facilitates interaction with other backend tools by allowing the agent to discover and trigger registered functions through the iii CLI.- [REMOTE_CODE_EXECUTION]: Headless turns are configured to auto-approve actions (always_approve: true), removing human-in-the-loop oversight for potentially dangerous shell operations.- [PROMPT_INJECTION]: Exposed to indirect prompt injection where malicious instructions in the filesystem or user messages could be executed by the Grok CLI.
Ingestion points: Content from the directory specified in cwd and input provided in prompt or messages.
Boundary markers: No explicit markers are used to differentiate untrusted data from the agent's instructions.
Capability inventory: Includes shell command execution, file system modifications, and cross-function triggering via the iii bus.
Sanitization: No input sanitization or validation is implemented before data is passed to the underlying execution engines.

grok