openclaw-behavior-plan

Generates structured behavior plans for OpenClaw agents based on user requirements. Use when the user asks to create a plan, design agent behavior, plan mult...

This skill appears to do what it says — it generates step-by-step plans that may reference tools which can read files or run shell commands. Before installing or invoking it, decide whether you will allow plans generated by the skill to access local files or execute commands. If you expect the agent to interact with calendars, Slack, or other services, grant those integrations only through the official connectors and avoid pasting secrets directly into prompts. Require explicit user confirmation before allowing any plan step that modifies or deletes system files or sends sensitive data to external endpoints.

Type: OpenClaw Skill Name: openclaw-behavior-plan Version: 1.0.0 The skill is designed to generate behavior plans for an AI agent, which can include highly sensitive operations. The `SKILL.md` explicitly instructs the agent to propose `execute_shell` for tasks like '删除、修改系统' (delete, modify system), and `examples.md` demonstrates the use of `read_file` to '识别潜在敏感信息（密钥、密码、token 等）' from files like `config.yaml`. While the skill itself does not execute these actions, it is designed to *plan* them based on user input, making it highly susceptible to prompt injection. A malicious user could craft a prompt to instruct the agent to generate a plan to read sensitive files (e.g., `/etc/passwd`, `~/.ssh/id_rsa`) and then, following its own instructions, identify and potentially output sensitive information, or to generate a plan for arbitrary shell command execution, leading to potential data exfiltration or remote code execution if the generated plan is executed.