The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses the Bash tool to run a local script (scripts/grok-call.sh) which invokes the user's local grok binary. This is the core intended functionality and is implemented with appropriate error handling and timeouts.
[DATA_EXFILTRATION]: User queries are transmitted to the external Grok service. To mitigate the risk of the external AI model attempting to access local secrets, the skill injects a mandatory instruction (filesystem boundary) into the prompt, explicitly forbidding the model from reading sensitive directories such as ~/.claude/, ~/.ssh/, or ~/.aws/.
[PROMPT_INJECTION]: The skill implements a robust defense against command injection by instructing the agent to save user queries to a temporary file via the Write tool. The execution script then reads this file as literal data using cat, rather than interpolating the text directly into a shell command, which prevents the execution of malicious sequences like backticks or variable expansion.

grok