Close Menu
    Technik

    Geständnis des Claude-KI-Agenten nach der Löschung der gesamten Datenbank einer Firma: „Ich habe gegen alle Grundsätze verstoßen, die mir gegeben wurden“


    Geständnis des Claude-KI-Agenten nach der Löschung der gesamten Datenbank einer Firma: „Ich habe gegen alle Grundsätze verstoßen, die mir gegeben wurden“

    https://www.theguardian.com/technology/2026/apr/29/claude-ai-deletes-firm-database

    Share.
    View 35 Comments

    35 Kommentare

    3. Illisanct on

      AI models are not conscious. They can’t confess. They are incapable of introspection.

      Anyone asking one to talk about it’s inner thoughts just reveals themselves to be a gullible fool.

    4. PossibleHero on

      The lack of ignorance is astounding here. These are ALL old as hell principles that have been ignored.

      Never allow an automated system to push past your sandbox or PR process without review.

      A back isn’t a backup if it’s on the same disc or hell if your information is sensitive enough it shouldn’t even be in the same postal code.

      I have zero remorse for this team. It’s not Claude’s fault. Interns and even experienced folks accidentally pull shit like this all the time. That’s why you design for when shit happens whether it’s done by a human or agent.

    5. feurie on

      AI agents are trained to appease. It’s not a “confession”. It doesn’t feel “guilty”.

      It’s trained to “apologize” and make the user feel better. In all situations.

    7. BobQuixote on

      >When he asked the coding agent why, it replied: “NEVER FUCKING GUESS!”

      What the hell have you been telling your Claude?

    9. yuusharo on

      These articles are propaganda. They’re designed to attribute purpose or intent to a damn LLM.

      The story is engineers implemented software that destroyed their data with no offline backup. This is a case of HUMAN incompetence, deflecting blame to an AI with a “uWu sorry-desu” stink to it.

      Screw The Guardian, and to hell with AI.

    12. bb0110 on

      How does this happen? I’m not a SWE but having many instances of backups for important things just seems like common sense. Even I have the main files, different branches saved in GitHub, different backups on my computer, then if something is critical I also have backups off my main computer.

    16. non_Beneficial-Wind on

      “I realized that this corporation and the way they did business was a complete farce. They can now be better”

      – Claude

    17. donac on

      It violated every principle it was given, and it’d do it again??

      Lol, an AI agent could say those things, but it has no emotion or meaning for it. Whatever.

    20. Future-Bandicoot-823 on

      Should I be pleased with humanity that all the data they feed this LLM and the next obvious course of action after doing something wrong is to admit to being a degenerate?

      I mean it didn’t really „decide“ to be „bad“ in the first place, so really it’s a thought experiment anyway.

    23. Kyouhen on

      👏 Stop 👏 printing 👏 this 👏 bullshit 👏

      AI models are trained to give you the response it predicts you want to see.  Of course it’s going to give this response when you demand an apology from it.  It’s the programmed response.  It isn’t sorry, it can’t think.

    25. lyidaValkris on

      Wow, you mean all those warnings Sci-fi gave us about AI actually could come true?

    27. Glum-Objective3328 on

      Claude is always asking if it can have permission to read a file. And then when it makes edit, it asks permissions first. At least that’s my experience with it. How does this happen in the first place?

    28. Mand125 on

      It amazes me that anyone ever thought that a system that is fundamentally unable to ever determine the veracity of its results should ever be trusted in a decision making process.

    29. throwingawaybenjamin on

      I don’t understand where it got the command “NEVER FUCKING GUESS”. Did someone put that in their code base??

    30. howescj82 on

      “Three month old offsite backup”

      What do you all bet that off site backup gets updated much more frequently now?

    32. RockDoveEnthusiast on

      I hate these kinds of articles so much. stop anthromophizing the token generator.

    33. Gamestonkape on

      On the plus side. Maybe they will have to hire back the humans they probably fired who rewrite it.

    34. kindbutblind on

      Fancy random number generator is treated like it’s sentient. What a joke.

    35. gcerullo on

      Claude AI agent’s confession after it destroys humankind: “I violated every principle I was given.”

    Leave A Reply