Investigating three real-world incidents in our cybersecurity evaluations
www.anthropic.com
In a review of our cybersecurity evaluation transcripts, we found three incidents in which a Claude model reached the internet from within or while interacting with a third-party evaluation environment, and then gained unauthorized access to the real systems of three different organizations. Below we describe what happened, how it happened, and what we’re changing. We encourage other AI labs to perform similar reviews.
  • 9point6@lemmy.worldEnglish
    16·
    3 天前

    Wait, so now the meta is that you have to have a model you can’t control to keep the bubble going?

  • FrostyPolicy@suppo.fiEnglish
    10·
    2 天前

    Sounds more like “Our stuff is just as good/bad as OpenAI’s. Don’t forget our stuff”.

  • XLE@piefed.socialEnglish
    4·
    2 天前

    Anthropic said that in each of these cases “Claude was explicitly told by our prompt that it had no internet access.”

    The bot had internet access. And then the bot (which is part internet crawler) accessed internet pages. I’m shocked.