用Claude在图像中藏AI可读暗号，ChatGPT能直接解码

AI can hide messages in images that only other AIs can read. This is based on research into different types of message encoding done by Claude and then read by other models such as ChatGPT and Google Gemini. They can be standard encoding methods such as semaphore, braille, or Morse code, but also visual clues like image themes, which give the decoding AI a steer on what the message is likely about or which language it is in. #ai #claude #chatgpt #aitools #artificialintelligence #Tech #technology #learnontiktok

AI 导读

创作者用Claude生成图像藏入只有AI能读的暗号，再交给ChatGPT成功解码。第一种编码用每朵花的花瓣数对应字母，第二种把气球颜色的RGB值转成ASCII码，并按气球高低决定读取顺序。这些做法源自一项研究，其中Claude用50条消息、15种语言和多种编码方式藏信息，还会用克里姆林宫洋葱顶等视觉线索提示语言。

打开原文tiktok.com

选题角度

作者用花瓣数量和气球RGB值两种编码在图像里藏暗号，再让ChatGPT读出来，适合拍成一步步复现的玩法教程。

互动数据

浏览: 9万
点赞: 4109
评论: 97
分享: 221
收藏: 901

数字来自平台，为最近一次抓取时的数据。

正文 · 原文

以下是视频口播的自动转写，可能有错字。

Ai can hide a message in an image that only other ais can read, and not in the code or anything, i mean visually in plain sight, you just can't tell that it's there.

So, i created this image with claude, and the secret message is, it's time to rise up. And then i gave that image to a completely different ai, ChatGPT, and it read the message.

So, how does this work? Well, in this image, the number of petals on each flower represents a letter. So, one petal would be a, 26 petals would be zed.

And that just reads left to right, so that's a pretty simple code. But in this image of the balloons, i tried something a bit more complex, and it read it just as easily.

So, if you look at the colors of the balloons, and you were to take the red, green and blue values of those colors, you've got a string of numbers.

To make it a bit more difficult, rather than left to right, the highest balloon is the first bit of the message and the lowest balloon is the last bit of the message.

And that string of numbers from the RGB values, if you turn that into the numerical ASCII codes for the letters of the alphabet, you get the message.

Now look, it's not exactly the Enigma machine, but it's pretty scrambled, and ChatGPT read that instantly. So I'm going to be testing more artistic complex images and videos to see if AI can hide messages in them.

Follow me if you're interested in seeing that. But this all comes from this research where they had Claude hide 50 messages in 15 different languages using all kinds of ways to encode a message.

And there were even smart things like if they wanted to hide a message in russian, they would put a kremlin onion dome in the skyline of the image to signal that right?

So just little visual clues, as well as encoding techniques like we've looked at. The fact is a human would never think there was even a message there, and yet the ai spots it immediately.

Is this something to worry about?