The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The "Remote Script Execution" template uses ssh "$remote_server" "bash -s" < "$remote_script" to execute local script content on a remote system, providing a direct vector for remote code execution.
[PRIVILEGE_ESCALATION]: Multiple templates, including those for package installation and service management, explicitly use sudo to acquire administrative privileges.
[PERSISTENCE]: The "Task Scheduler (Cron Setup)" script automates the modification of the system crontab to ensure scripts run automatically at scheduled intervals, which is a standard method for establishing persistence.
[CREDENTIALS_UNSAFE]: The "Database Backup Script" template contains hardcoded variables for database credentials (db_user and db_pass) and passes them as plain-text command-line arguments to mysqldump, exposing them to other users via the system process list.
[COMMAND_EXECUTION]: The skill generates shell scripts that perform sensitive or destructive actions (such as rm, useradd, and systemctl) based on variable interpolation.
[INDIRECT_PROMPT_INJECTION]: The templates ingest untrusted data like file paths, server addresses, and usernames directly into shell commands without sanitization, creating an attack surface for command injection if the agent interpolates malicious strings into these variables.
Ingestion points: Variables such as $backup_dir, $source_dir, $remote_server, $database_name, and $username in SKILL.md.
Boundary markers: Absent.
Capability inventory: File deletion (rm), remote execution (ssh), privilege escalation (sudo), and persistence management (crontab).
Sanitization: Absent.

Linux Production Shell Scripts