The Agent Skills Directory

[COMMAND_EXECUTION]: The skill relies on the Bash tool to execute a series of shell commands. It specifically uses ssh to run diagnostics on remote servers, which is an inherently high-privilege activity.
[DATA_EXFILTRATION]: The diagnostic logic includes reading process environment variables via cat /proc/{PID}/environ. This file frequently contains sensitive data such as API tokens, private keys, or internal credentials passed to the process at runtime.
[COMMAND_EXECUTION]: The skill is vulnerable to command injection through unsafe interpolation. It takes values derived from remote command output (such as {CONTAINER} names and {PID} values) and user-supplied strings ({SSH_TARGET}) and inserts them directly into shell command strings without validation or escaping. This allows an attacker to execute arbitrary commands if they can control the server output or the target string.
[EXTERNAL_DOWNLOADS]: The skill targets a specific hardcoded IP address 124.158.103.16 for SSH operations. While functional, this represents a network connection to an external endpoint not managed by the agent platform.
[PROMPT_INJECTION]: The skill lacks boundary markers or sanitization logic when processing data from the remote server. Because it trusts the output of commands like docker ps to form subsequent commands, it is susceptible to indirect prompt injection if the server returns malicious metadata designed to subvert the agent's instructions.

gpu-use