The Agent Skills Directory

[COMMAND_EXECUTION]: The skill is vulnerable to arbitrary code execution because it directly interpolates user input from $ARGUMENTS into a Python script string executed via python3 -c without any sanitization or escaping.
Evidence: In SKILL.md, the line filt = '$FILTER'.strip().lower() allows an attacker to 'break out' of the Python string literal and execute arbitrary commands by providing an input such as '; import os; os.system("id"); #.
[REMOTE_CODE_EXECUTION]: The injection vulnerability in the Python processing logic provides a direct vector for remote code execution on the host machine if the agent passes malicious strings from untrusted sources into the arguments.
[CREDENTIALS_UNSAFE]: The script reads the Clash Verge API secret from local configuration files and prints it directly to the terminal, exposing a sensitive credential to the agent's context and system logs.
Evidence: In SKILL.md, the line echo "Secret: ${SECRET:-(未找到)}" explicitly prints the extracted authentication secret.

clash-routes