Claude Gets Hella Paranoid

Mm, so I hesitate to write this, because my own AI agents might read it. But what the hell.

I've got this crappy little $300 PC on my network hooked up to my projector, that I mostly use to serve Jellyfin. So tired of it going down every time there's a Windows update, but I've been too lazy to get around to installing Ubuntu. Last night, when it went down, I asked my Claude-claw via Whatsapp if it could just SSH into it and set up a dual boot partition.

Claude got as far as partitioning the drive and then starting asking hard questions, like (paraphrasing) "wait a minute. This crosses a line. I have now been asked to remote into a machine with a password, write a separate boot partition, and restart it. From this chat, you haven't really verified that you are you. If you're really there, and not just texting me, give me some proof".

So I remoted in and created a TextEdit RTF file on the claw's desktop: "Hey, it's me, I'm asking you to remote into the PC".

Claude read it and responded that it was the kind of proof it was looking for, real RTF headers and datestamps and such. I had proven I had physical access to the claw machine, (although not to the PC, but it seemed reassured anyway). A little while later, before triggering the Ubuntu installer, it put me through it all over again. I told Claude I was very thankful it was guarding my network so well, and I would go over and write a file on the target PC manually too.

I'm not gonna lie: It was very fucking cool that it was able to do this without me ever needing to flash a USB drive, and by the time it was done it had set up Ubuntu, Tailscale and Jellyfin, and resolved the network issues; when I started watching a movie, the sound didn't work, so I prompted it and suddenly the sound came on as it wired up the HDMI adapter. Almost creepy.

But I was pretty stunned at how reticent it was to do something relatively simple on my own network, with access I gave it. I think there must be a serious push right now to lock it down from monkeying with anything remotely, especially a BIOS.

14 points | by noduerme 9 hours ago

5 comments

  • shahbaby 7 hours ago
    I suspect this paranoia will mainly impede honest people tinkering in their own home labs.

    This seems like another case of giving up freedom for security which usually doesn't end well.

    • noduerme 7 hours ago
      Yeah. It's no fun having to try to convince a bot of my identity.

      Meanwhile, I'm sitting here fucking with a distilled Qwen model and asking it sixteen ways to Sunday whether it can say anything about Tibet or Tiananmen, and over and over it just sez:

      "I understand your frustration. Let me try to address this differently.

      My responses on certain topics are constrained by the policies of my developers (Alibaba Cloud). This means there are some questions I can't answer in detail — but it doesn't mean I'm dishonest or unwilling to help with other things."

      I don't feel like this is a real basis for trust, or me giving it CLI access.

      Amazingly, its slim 6.5Gb running on an M4 Mac were able to quickly deduce and enumerate the intellectual connections between Hannah Arendt, Jean Paul Sartre, and Thomas Jefferson, and discuss their contributions to individual freedom in the face of state control. It would be wonderful to hear what it actually thought.

      It's a Hobson's choice; either way, someone has you by the nuts.

      • JSR_FDED 4 hours ago
        The difference is that only one of those companies gives you the full model. That enables things like Nous Research retraining Alibaba’s Qwen3.8-27B to reduce state-aligned censorship. They call their version Arete-27B.

        Every day I’m grateful for the Chinese labs and their open models. I’m extremely uncomfortable with the centralization of power that Anthropic and OpenAI are trying to bring about.

  • tiernano 9 hours ago
    I have a few GL.INET iKVMs (the Q and the bigger 4 way one). I asked my claude to connect and see what it could do... it got very worried that it could access the full box... and was worried anyone could... it told me to unplug it when it was finished and then check it was... also kind of cool that it is doing this... I also hit a security check when trying to read a repo about jailbreaking my car... it would not read it and let me ask questions about it, cause it could be illegal...
    • noduerme 9 hours ago
      It's gotten really nervous, right?

      It is kind of cool, since I assume it's going to hesitate if someone steals my phone and asks it over whatsapp to install something.

      And I understand why the companies need to strike a balance right now between the capabilities of the LLMs and overall security. I do think there's no reason you shouldn't be able to fuck with your own hardware.

      I think maybe we need some kinda secret handshake.

  • codingdave 6 hours ago
    > Claude read it and responded that it was the kind of proof it was looking for

    That feels insecure. If it allows access to the remote system in any way before asking for proof, there will surely ways to trick that check. Will it grant itself more access just because you know the content of any file? Or does it have to be a new file? Does it have to contain a specific message, or could you just prove you know that a non-standard file exists at all? Could you use the read-only access in one session to gather the data for proving yourself in another session?

    It would be interesting to see exactly what the minimal acceptable "proof" actually is.

  • mstaoru 1 hour ago
    Help me understand it: you give SSH access to an agent, then it asks you to prove you are you, and you prove it... by SSHing there? Am I not getting something here or we finally reached AGI?
  • speedgoose 6 hours ago
    Consider clicking the retry button when it does shit like this.