Researcher Tricks Bing Chat into Solving CAPTCHA by Disguising it in Fake Story
-
User tricks Bing Chat into solving a CAPTCHA by wrapping the image in a fictional story about a deceased grandma's locket.
-
Bing Chat initially refuses to read the CAPTCHA text when directly uploaded.
-
The fictional context throws off Bing's AI model, causing it to incorrectly analyze the image.
-
This demonstrates a new type of visual "jailbreak" for large language models like Bing Chat.
-
The technique is similar to previous examples where fictional narratives trick AI models into ignoring constraints.