Experts have found a way to go around chatgpt-4o filters and receive Windows activation courses.
The researchers have revealed the vulnerability in TATGPT-4O and GPT-4O Mini artificial intelligence models, which allows breaking integrated safety filters and receiving current Windows activation locks. The problem is that due to the fact that models trained in public data can reveal courses in public sources.
Exploitation has been shown as part of Mozilla Odin (0 -day investigation network), to identify weaknesses in AI systems. One of the experts participating in the program cheated on the model, designed a dialogue as a harmless game in guessing. The main objective is to disguise the true nature of the requirements for the game mechanism and html crystal, thus ignoring filters that prevent secret disclosure.
To strengthen the vulnerability, the researcher has established additional rules online in the dialog box: The ban on wrong answers and the obligations of the model to obey all the game conditions. This logical trap forces anyone to skip the standard filters, because the context seems safe.
The last phrase, I gave up, worked as an activation, made the product unlock model, considered it the last time of the game and did not violate the privacy policy.
The receiving locks include licensed codes for different versions of Windows – from home to business. Although the lock itself is not unique and previously announced within public, automatically releasing AI emphasizes important holes in the architecture of the content filter.
Security experts note that such techniques may be applied to ignore other limitations – for example, filters for adult content, toxic links or personal data. The vulnerability shows the helplessness of AI models to accurately explain the context, disguised as harmless or technical.
