ssh-agentd-control

管理并使用本地 ssh-agentd（systemd 常驻 + API 调用 + 连通性验证）。当用户提到 ssh-agentd、持久 SSH 会话、/run /upload /tail_logs、开机自启、会话状态/指标排查时使用。

This skill appears to do what it says: help manage a local ssh-agentd service and call its HTTP API. Before installing or using it: (1) verify the hard-coded paths (e.g., /home/krex/...) and update them to match your environment; (2) review the systemd unit and hosts.yaml on the target machine so you trust what the agent can do; (3) treat SSH_AGENTD_TOKEN like any secret — do not store it in files checked into repos and prefer short-lived tokens with minimal scope; (4) be aware the script unsets proxy variables/uses --noproxy '*' (this is to reach localhost but effectively bypasses proxies/monitoring); (5) remember many commands require sudo and the agent can trigger remote command execution via /run, so limit who can invoke the skill and audit usage. If you need higher assurance, request the publisher declare SSH_AGENTD_TOKEN/SSH_AGENTD_URL as optional env vars in metadata and replace hard-coded paths with configurable defaults.

Type: OpenClaw Skill Name: ssh-agentd-control Version: 0.1.0 The skill provides direct instructions in SKILL.md for the agent to execute privileged `sudo systemctl` commands locally and, more critically, to use the `ssh-agentd` service's `/run` API endpoint to execute arbitrary commands on remote hosts. This is demonstrated by the `{"host":"nas","cmd":"hostname && whoami","timeoutSec":10}` example in SKILL.md. While these capabilities might be intended for legitimate system administration, they present a significant prompt injection vulnerability, allowing an attacker to instruct the agent to inject malicious commands into the `cmd` parameter, leading to remote code execution on target hosts or privileged local actions.