I inspected Claude Code for privacy reasons and found hidden system prompt markers based on API base URL and timezone.
  • Strider@lemmy.worldEnglish
    17 days

    I can just shake my head and say that humanity won’t make it. We are so obsessed with stealing things from others but making sure our offer is not stolen it’s insane.

    Instead of really working together, which we are in dire need of currently.

  • jungle@lemmy.worldEnglish
    17 days

    What a stupid take, and entirely AI written to boot.

    Of course Anthropic wants to detect if they are getting distilled.

    And of course they need to do that covertly. Does the “author” of this article expect them to tell attackers how to hide? Really?

    • 4am@lemmy.zipEnglish
      17 days

      When did nerds get so spoiled and entitled? Was it when every parent on America sent their little snot nosed video game playing asshole to college so he could “be the next bill gates”?

      The current “AI boom” is the first settler-colonial project of knowledge and it must be treated as such.

    • mal3oon@lemmy.worldEnglish
      17 days
      • Step 1: Train a model on all the stolen content on the internet you can get your hands on
      • Step 2: Complain and bitch about all the other LLMs paying you money to use your model to train theirs, and try to prevent it
      • Step 3: ??
      • Step 4: Profit
      • redjard@reddthat.comEnglish
        16 days

        Ah but you see in the agreement it says the llm tex output cannot be used to train your ai.
        You have to respect the license that text is published under when you train your ai on the internet, as long as the internet is a website by anthropic.

  • frongt@lemmy.zipEnglish
    17 days

    That’s not what steganography is. There is no hidden message here.

    • WesternInfidels@feddit.onlineEnglish
      17 days

      Your assertion baffles me. The CC client is sending information about its execution context back to Anthropic HQ in a sneaky, obfuscated way that most people wouldn’t notice.

      If that’s not a hidden message, if that’s not steganography, what is?

      • frongt@lemmy.zipEnglish
        16 days

        It’s not sending it back. It’s essentially watermarking the output, so that it can be identified if it appears somewhere unexpected.

        • QuizzaciousOtter@lemmy.dbzer0.comEnglish
          16 days

          Did you read the post?

          the marker becomes part of the system context sent to the model. (Where Anthropic probably parses in their backend)

        • GamingChairModel@lemmy.worldEnglish
          16 days

          It modifies the prompt, aka the input, not the output. It is smuggling 3 bits of secret user/session data in a wrapper that doesn’t look like it contains that data. As the article explains:

          So the marker becomes part of the system context sent to the model.

          This is a normal timestamp on a prompt:

          Today's date is 2026-07-11.

          But if your system timezone is a Chinese mainland timezone, it looks like:

          Today's date is 2026/07/11.

          Then, if your base URL includes a keyword like “deepseek,” it silently replaces the apostrophe from a ' to a ʼ:

          Todayʼs date is 2026-07-11.

          Or if the base URL has one of the domains on the list, like any .cn domain, it replaces the apostrophe with another apostrophe character:

          Today’s date is 2026-07-11.

          And if it has both a URL and a keyword on the watchlist, the prompt context includes:

          Todayʹs date is 2026-07-11

          That’s 3 bits of information: does this system have a mainland Chinese time zone, does the base URL contain a known keyword (associated with Chinese AI competitors) or a known domain (associated with mainland China or its major tech companies). And it sneaks it on by without making it obvious.

          That’s steganography.