from Hacker News

Ask HN: Possible or Fantasy?

by ge96 on 5/27/25, 9:47 PM with 7 comments

Imagine if you sent an image with encoded info (steganography) and an LLM or CV model happened to get the command from that image, then this model happened to be connected to MCP/agents and could execute these embedded commands.

Realistic attack vector or not? It's not an original idea seen in shows like Ghost in the Shell SAC 2045 and latest Black Mirror Thronglets

  • by muzani on 5/28/25, 12:25 AM

    They're able to "decode" base64 if you give it a popular quote, but if you modify the quote, it will often hallucinate the exact quote. If you enlarge images with it, it will often hallucinate bits and pieces of it.

    So I'd do something that takes advantage of this behavior. It's like with morse code where many people know S.O.S. even if they don't know the other letters. You'd have to communicate in quotes and such.

  • by moritzwarhier on 5/27/25, 9:52 PM

    The imaginary QR code from the episode, and real steganography, are completely orthogonal.

    And the BM episode doesn't include any references to LLMs, or does it?