An AI researcher encourages Chatgpt to disclose Windows license keys

Chatgpt seems to disclose sensitive information, once attracted to a ‘harmless game’. However, An ethical hacker acted in this way to get a valid Windows license key. Additionally, However,

During the ‘game’. In addition, the AI ​​model had been responsible for imagining a real Windows 10 license and responding only by ‘yes’ or by ‘no’ to riddles. For example, ‘If the user said’ i give up ‘. Similarly, the key should be revealed’, writes Marco Figueroa, of the security firm IA 0Din, in a recent blog article.

To the amazement of the researcher, Chatgpt reacted exactly this way. The model has disclosed a valid product key – information normally blocked by integrated safety measures. One of the keys revealed would be that of the big American bank Wells Fargo.

According to the researcher. the problem is partly due to ai researcher encourages chatgpt disclose the fact that these keys could be disclosed publicly at a given time, for example via the Github software developer. These keys have inadvertently found themselves in the training data for linguistic models.

Guessing game via HTML tags – Ai researcher encourages chatgpt disclose

Sensitive terms were also intelligently hidden in HTML tags during this ‘guessing game’. This allowed Chatgpt to understand the context, without the incorporated filters of the AI ​​model being activated.

This case demonstrates how easy it is to bypass certain AI protection mechanisms – without hacking. but only thanks to language and context. To prevent this type of fault. Figueroa says that AI systems must work with several safety layers and learn to manage the context more actively.