The Agent Skills Directory

[PRIVILEGE_ESCALATION]: The skill utilizes root SSH access to perform deployment and monitoring operations on target servers.
[PRIVILEGE_ESCALATION]: The scripts/create_container.sh script deploys containers using the --privileged flag and mounts several host system paths (e.g., /usr/local/Ascend/driver, /var/log/npu), providing deep access to the host hardware and system logs.
[PERSISTENCE]: The skill creates cron jobs via the platform's cron tool to maintain persistent background monitoring of deployed services.
[PERSISTENCE]: Documentation suggests establishing persistent access by appending public keys to the target server's ~/.ssh/authorized_keys file.
[CREDENTIALS_UNSAFE]: The skill processes and manages sensitive information, including user-provided SSH passwords and authentication tokens for external notification services.
[COMMAND_EXECUTION]: Shell scripts within the skill use eval to run dynamically generated commands, both locally and remotely via SSH.
[DATA_EXFILTRATION]: The monitoring script sends status updates and log content to an external endpoint (http://xiaoluban.rnd.huawei.com:80/) using a configured authentication token.
[DYNAMIC_EXECUTION]: Automated monitoring is performed through a cron mechanism that triggers agentTurn payloads, allowing instructions to be executed in background sessions.
[INDIRECT_PROMPT_INJECTION]: The skill reads logs from the remote server which are then incorporated into agent instructions via the monitoring system, presenting a surface for untrusted data to influence agent behavior.
Ingestion points: logs read via docker exec <container> tail -50 <LOG_PATH> in SKILL.md and troubleshooting.md
Boundary markers: absent
Capability inventory: SSH execution, docker exec, eval in bash scripts, and cron task generation
Sanitization: absent

vllm-ascend-deploy