Safety安全说明
Understanding Agent Safety Risks了解智能体安全风险
When an agent can operate a phone, it may read what appears on screen, tap buttons, open apps, and trigger workflows. That is useful, but it also means mistakes or misuse can have real consequences.当智能体可以操作手机时，它就可能读取屏幕内容、点击按钮、打开应用并触发工作流。这很有用，但也意味着一旦出错或被滥用，就可能产生真实后果。
1. Account and privacy risk1. 账号与隐私风险
If the device is logged into chat tools, social apps, or cloud services, an agent may accidentally expose sensitive content, send the wrong message, or interact with the wrong account.如果设备已经登录了聊天工具、社交应用或云服务，智能体就可能误读敏感内容、发错消息，或者在错误的账号下执行操作。
2. Permission risk2. 权限风险
Accessibility, screen reading, notification access, and similar permissions are powerful. If you enable them casually, you may give the app more control than you intended.无障碍、读屏、通知访问等权限都非常强大。如果你在没有充分理解的情况下开启这些权限，可能会让应用获得超出你预期的控制能力。
3. Automation mistakes3. 自动化误操作
Even a well-designed agent can make mistakes. It may tap the wrong UI element, misunderstand a screen state, repeat an action, or continue a workflow longer than you expected.即使是设计良好的智能体，也可能出现误操作。它可能点错控件、误判页面状态、重复执行动作，或者比你预期中运行得更久。
4. External service and content risk4. 外部服务与内容风险
If the agent interacts with websites, messages, or third-party services, it may encounter malicious content, fraudulent prompts, or unreliable information. A device-side agent is not automatically safe just because it runs locally.如果智能体会接触网页、消息或第三方服务，它也可能遇到恶意内容、欺骗性提示或不可靠信息。即使在本地设备上运行，也不代表天然安全。
How to reduce risk如何降低风险
- Use a dedicated secondary phone instead of your primary device.优先使用备用设备，而不是主力手机。
- Use separate app accounts for testing and daily automation.使用独立账号来进行测试与日常自动化。
- Enable only the permissions that are truly required.只开启真正需要的权限。
- Review workflows carefully before letting the agent run unattended.在让智能体脱离人工监督前，先仔细检查工作流。
- Avoid spam, abuse, or any behavior that could affect other people or services.避免任何骚扰、滥用或可能影响他人和服务的行为。
A good default mindset一个稳妥的默认心态
Treat MobileClaw like a powerful experimental tool, not a harmless toy. Start small, keep the environment controlled, and expand only after you understand what the agent can do on your device.可以把 MobileClaw 当作一个强大的实验性工具，而不是无害的小玩具。先从小范围、受控环境开始，确认你真正理解它在设备上能做什么，再逐步扩大使用范围。