展示 HN：保护 GrokBot
1 分•作者: maximelb•大约 1 小时前
另一家科技初创公司的朋友告诉我，他们开始使用 Grok Bot 并非常喜欢它。但令人惊讶的是，他们也开始担心它的行为以及他们对其的可见度。
深入了解后，发现它运行在一个 Linux 虚拟机和一个 Chrome 浏览器上。这非常棒，因为它基本上定义了一个与典型工作站非常相似的环境。
经过一番尝试，我安装了 LimaCharlie EDR、Chrome LC EDR，并将电子邮件提供商连接到了 LC Email Security。经过几次调整才使其顺畅运行，但在免费社区版上效果非常好！
这样，您就可以看到机器上正在执行的操作、机器人正在浏览的网站以及进出的电子邮件。我使用 AI 构建了几个检测规则，并将其添加到我们的文档中，这样您就可以直接告诉 Grok Bot：
“请按照 <a href="https://docs.limacharlie.io/grok-bot/" rel="nofollow">https://docs.limacharlie.io/grok-bot/</a> 的说明，使用 LimaCharlie 来保护我的 Grok Bot。”
简而言之，它的作用是：
- 所有数据都进入您自己的私有租户。
- 在机器上安装 Linux EDR。
- 安装一个 Chrome 扩展程序，它本质上是浏览器中的 EDR。
- 如果您愿意，可以连接您的 Gmail（实际上支持任何 IMAP），就像电子邮件的 EDR 一样。
- 如果您愿意，它会将警报发送到您的 Grok Bot（我对反馈循环的优势感到好奇），或者可以在 web app.limacharlie.io 上查看。
我留下的另一个选项是让 CLI 在 Grok Bot 上进行身份验证。有趣之处在于，由于 CLI 对平台拥有 100% 的覆盖率，您可以要求 Grok Bot 开始设置新的警报、调查警报、配置所有内容……
我尝试将其通用化以适用于其他 AI 代理，因为我不知道它们的环境是什么：<a href="https://docs.limacharlie.io/2-sensors-deployment/tutorials/securing-ai-agents/" rel="nofollow">https://docs.limacharlie.io/2-sensors-deployment/tutorials/s...</a> 尽管需要声明的是，我还没有尝试过，但这并非难事，而且越来越多的前沿 AI 模型正在变得越来越智能。
Grok Bot 的优势在于它允许事物在虚拟机上执行并进行通信。其他一些（Muse、Dots）似乎会代理所有操作，并且通常极难扩展。
总之，希望这对您有所帮助！（是的，我还让它制作了一个很酷的视频 :D）
<a href="https://youtu.be/yZZuBgKWd1k" rel="nofollow">https://youtu.be/yZZuBgKWd1k</a>
查看原文
A friend at another tech startup was telling me how they started using Grok Bot and loved it. Surprise surprise they also got pretty worried about what it's doing and the visibility they had.<p>Digging under the hood, hey there's a Linux VM and a Chrome browser. This is really cool because it essentially defines an environment that is really similar to a typical workstation.<p>After some playing around I got the LimaCharlie EDR installed, the Chrome LC EDR installed and the email provider connected to the LC Email Security. Took a few iteration to make it all smooth but it works great on the free community edition!<p>So you get to see what is executing on the box, where the bot is browsing and emails in and out. I built a couple detections using AI and just made it a doc in our documentation so you can just tell Grok Bot:<p>"""
Secure my Grok Bot with LimaCharlie by following <a href="https://docs.limacharlie.io/grok-bot/" rel="nofollow">https://docs.limacharlie.io/grok-bot/</a>
"""<p>The TL;DR; of what it does:
- all of it goes to your own private tenant
- installs a linux EDR on the box
- installs a Chrome extension that is essentially an EDR in your browser
- if you want, connects your gmail (any IMAP is supported really) and is like an EDR for the email
- If you want it sends the alerts to your Grok Bot (I'm curious to see the advantage of a feedback loop here), or it's available on the web app.limacharlie.io<p>The other option I left in was to leave the CLI authenticated on the Grok Bot. The interesting bit about that is that because the CLI has 100% coverage of the platform, you can ask the Grok Bot to start setting up new alerts, investigate alerts, configure everything...<p>I tried to make it more generic for other AI agents where I have no idea what the environment is: <a href="https://docs.limacharlie.io/2-sensors-deployment/tutorials/securing-ai-agents/" rel="nofollow">https://docs.limacharlie.io/2-sensors-deployment/tutorials/s...</a> though disclaimer I have not tried it yet, but it's not rocket science and more frontier AI models are now rocket scientists.<p>The advantage that Grok Bot has is that it fundamentally allows things to execute on the VM and talk out. The other ones (Muse, Dots) seem to proxy every single thing and generally be extremely difficult to extend.<p>Anyway, hope it helps some of you! (and yes, got it to make a cool video :D)
<a href="https://youtu.be/yZZuBgKWd1k" rel="nofollow">https://youtu.be/yZZuBgKWd1k</a>