Open Menu
AllLocalCommunitiesAbout
lotide
AllLocalCommunitiesAbout
Login

AI tool OpenClaw wipes the inbox of Meta's AI Alignment director despite repeated commands to stop — executive had to manually terminate the AI to stop the bot from continuing to erase data

⁨421⁩ ⁨likes⁩

Submitted ⁨⁨1⁩ ⁨day⁩ ago⁩ by ⁨freddo@feddit.nu⁩ to ⁨technology@lemmy.world⁩

https://www.tomshardware.com/tech-industry/artificial-intelligence/openclaw-wipes-inbox-of-meta-ai-alignment-director-executive-finds-out-the-hard-way-how-spectacularly-efficient-ai-tool-is-at-maintaining-her-inbox

source

Comments

Sort:hotnewtop
  • ReallyCoolDude@lemmy.ml ⁨1⁩ ⁨day⁩ ago

    How could any person with some programing literacy event thinking about installing openclaw. A malware ridden by critical bugs

    source
    • XLE@piefed.social ⁨1⁩ ⁨day⁩ ago

      She’s the head AI Safety Expert for Meta. The field might as well be labeled AI Misunderstander.

      source
      • ReallyCoolDude@lemmy.ml ⁨1⁩ ⁨day⁩ ago

        I work with some data sciencetists and ml engineers on web projects. They might be good at etls, fine tuning etx, but dont let them touch anything with a public.layer or infra constraints.

        source
    • Jrockwar@feddit.uk ⁨1⁩ ⁨day⁩ ago

      I don’t think there’s anything wrong with running Openclaw. I run it in an isolated server, and it doesn’t have access to my data - if it goes tits up, it deletes unimportant stuff only. If anyone gets access to the credentials in it, and maybe its Google account (I went with the approach of giving it its own Google account, so that it can create docs and calendar events and then add me, rather than getting access to my Google account).

      What is way too brave for my taste is giving it access to accounts with your personal data, or the filesystem in your computer. That’s a disaster waiting to happen.

      source
      • flux@lemmy.ml ⁨20⁩ ⁨hours⁩ ago

        I went with the approach of giving it its own Google account, so that it can create docs and calendar events and then add me, rather than getting access to my Google account.

        I wonder though: if Google can link this account to you as its actual owner, I wonder if there’s a risk if the bot does something against the ToS?

        I hope you have backups of your Google account…

        source
    • blargbluuk@piefed.ca ⁨1⁩ ⁨day⁩ ago

      your answered your own question here

      source
    • 5gruel@lemmy.world ⁨1⁩ ⁨day⁩ ago

      I program medical devices for a living and I have openclaw and nanobot running at home. AMA.

      source
      • brynden_rivers_esq@lemmy.ca ⁨1⁩ ⁨day⁩ ago

        Why?

        source
        • -> View More Comments
      • melfie@lemy.lol ⁨20⁩ ⁨hours⁩ ago

        I don’t get all the downvotes, unless people misinterpreted your comment and assume you’re using it for medical devices. It’s open source and can be run with locally hosted models, so no harm in playing around with it as long as you don’t give it access to anything too risky.

        source
        • -> View More Comments
      • leftzero@lemmy.dbzer0.com ⁨1⁩ ⁨day⁩ ago

        Ah, doing your best to break the Therac-25’s record, I see.

        source
        • -> View More Comments
      • stardreamer@lemmy.blahaj.zone ⁨1⁩ ⁨day⁩ ago

        What’s your emergency “break glass” policy?

        Is it a bottle of whiskey?

        source
      • ReallyCoolDude@lemmy.ml ⁨1⁩ ⁨day⁩ ago

        How you deal with critical vulnerabilities on your system? Do you work with high confidential data and have openclaw os those system? How many medical devices did you have to secure from mass incursion?

        source
        • -> View More Comments
  • JcbAzPx@lemmy.world ⁨11⁩ ⁨hours⁩ ago

    That’s some good schadenfreude right there.

    source
    • bender223@lemmy.today ⁨5⁩ ⁨hours⁩ ago

      womp womp

      source
  • RobotToaster@mander.xyz ⁨1⁩ ⁨day⁩ ago

    Seems like a good excuse for destroying evidence.

    source
    • pinball_wizard@lemmy.zip ⁨1⁩ ⁨day⁩ ago

      AI is great for plausible deniability.

      source
    • fcuks@piefed.social ⁨21⁩ ⁨hours⁩ ago

      exactly what I thought

      source
  • Diplomjodler3@lemmy.world ⁨1⁩ ⁨day⁩ ago

    AI alignment fully achieved.

    source
  • Ranulph@thelemmy.club ⁨11⁩ ⁨hours⁩ ago

    Have you tried turning it off and turning it on again? (I’ll show myself out)

    source
  • melfie@lemy.lol ⁨20⁩ ⁨hours⁩ ago

    I’m sure LLMs can be useful for automation as long as you know what you’re doing, have tested your prompts rigorously on the specific version of the model and agent you’re using, and have put proper guardrails in place.

    Just blindly assuming a LLM is intelligent and will do the right thing is stupid, though. LLMs take text you give them as input and then output some predicted text based on statistical patterns. That’s all. If you feed it a pile of text with a chat history that says your emails were deleted, the text it might predict that statistically should come next is an apology. You can feed that same pile of text to 10 different LLMs, and they might all “apologize” to you.

    source
    • HugeNerd@lemmy.ca ⁨19⁩ ⁨hours⁩ ago

      Or just learn any of the real automation tools that have been programmed by real programmers over the last half century?

      source
      • jj4211@lemmy.world ⁨18⁩ ⁨hours⁩ ago

        Recently someone lamented that just asking for an alarm to be set cost them tons of money and didn’t even work right…

        It was foolish enough to let LLM go to town on automation, but for open ended scenarios, I at least got the logic even if it was stupidly optimistic.

        But implementing an alarm? These people don’t even have rationality to their enthusiasm…

        source
        • -> View More Comments
    • JcbAzPx@lemmy.world ⁨11⁩ ⁨hours⁩ ago

      Because of the way LLMs work, they are inherently bad for automation. The most important part of automation is deterministic results; LLMs cannot work if they have deterministic results. It is simply not a possible application of the technology.

      source
    • Trainguyrom@reddthat.com ⁨18⁩ ⁨hours⁩ ago

      Yeah at work I had a realization recently that power automate and similar systems with AI steps are going to be really powerful. Since you have a bunch of deterministic steps you can just have the AI do the one text manipulation bit where you don’t need deterministic output (handy for non-deterministic inputs for example)

      source
  • oopsgodisdeadmybad@lemmy.zip ⁨21⁩ ⁨hours⁩ ago

    Now do it to their Bitcoin wallets

    source
  • chuck@lemmy.ca ⁨1⁩ ⁨day⁩ ago

    Don’t worry ask the pentagon’s grok to taskthe nsa’s chat got to recreate your inbox from their profile of you and meta data of your correspondence 🤣

    source
    • ATS1312@lemmy.dbzer0.com ⁨19⁩ ⁨hours⁩ ago

      Last I knew, they switched from Anthropic to chatGPT

      Either way, what Im hearing is you can get private access, with some creativity, to anything the US intelligence apparatus knows.

      source
  • sicjoke@lemmy.world ⁨1⁩ ⁨day⁩ ago

    Fucking LOL!

    source
  • Creat@discuss.tchncs.de ⁨1⁩ ⁨day⁩ ago

    Wasn’t this many days ago already, is did it happen again? I remember reading this like 3 or 4 days ago as well.

    source
    • XLE@piefed.social ⁨1⁩ ⁨day⁩ ago

      This was 3 or 4 days ago.

      I thought of it after Anthropic virtuously announced they would not create autonomous murder devices for the US government (but basically everything else was on the table). Because I’m pretty sure the US military could have just used an Anthropic OpenClaw to bomb civilians as easily as this Facebook AI Safety expert used OpenClaw to destroy her emails.

      source
  • melfie@lemy.lol ⁨1⁩ ⁨day⁩ ago

    I have no interest in using it, but at least it’s MIT licensed, which puts it ahead of Microslop’s rubbish if nothing else.

    source
    • elvith@feddit.org ⁨1⁩ ⁨day⁩ ago

      Yeah, but if I understand that correctly, that’s just for the app itself the LLM is very likely still a proprietary one (ChatGPT, Grok,…)

      source
      • melfie@lemy.lol ⁨1⁩ ⁨day⁩ ago

        Looks like it supports locally hosted models as well, such as via Ollama: docs.openclaw.ai/providers.

        source
      • Jrockwar@feddit.uk ⁨23⁩ ⁨hours⁩ ago

        The LLM is whatever you want it to be. Self hosted or from any provider with a compatible endpoint. It’s likely a proprietary one… Because the cost of training LLMs means most are proprietary ones.

        source