The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructs the agent to execute arbitrary shell commands in other terminal windows via the /send and /create (autorun) API endpoints. This allows the agent to perform any action the user's shell is capable of, including file deletion or system modification.
[COMMAND_EXECUTION]: The instructions explicitly teach the agent how to bypass safety mechanisms in other AI tools (like Claude CLI) using flags such as --dangerously-skip-permissions or by simulating automated keypresses to bypass interactive confirmation prompts.
[DATA_EXFILTRATION]: The skill provides the ability to read the output history (up to 2000 lines) of any terminal window via the /read endpoint. This exposes potentially sensitive data, including environment variables, logs, and previous command outputs, to the agent's context.
[PROMPT_INJECTION]: The skill is highly susceptible to indirect prompt injection. It is designed to take natural language input ('User white speech') and translate it into executable shell commands or cron schedules. There are no instructions for sanitizing user input or using boundary markers to prevent malicious user requests from being executed directly.
[CREDENTIALS_UNSAFE]: The skill manages a sensitive FANBOX_CTL_TOKEN used to authenticate API requests. While the instructions warn against leaking the token to files, the agent must include it in all outgoing curl requests, making it accessible within the agent's execution history.

fanbox-agent