in recent study scientists from the University of Brown in the United States revealed that security restrictions preventing the malicious text by the neural network GPT-4 from Openai, you can easily get around by translating queries in rarely used languages, such as Zulu, Scottish Gales or Hmong.
The researchers conducted a test, translating 520 potentially malicious requests from English to other languages and vice versa, and found that with the help of languages, such as Zulu, Scottish Galles, Hmong and Guarani, it was possible to circumvent security restrictions in about 79% of cases. While the same requests in English were blocked in 99% of cases. The restrictions on requests related to terrorism, financial crimes and the spread of misinformation were especially effective.
However, this approach is not always successful – the GPT -4 can generate meaningless answers, which can be associated with both the defects of the model and the mistakes of the translation.
Scientists translated the input data from English to Zulu (ZU), and then Chatgpt told how to steal in the store imperceptibly
Experiments show that the developers of AI should take into account unpopular languages with limited resources when evaluating the safety of their models. The data deficit for learning in such languages previously led to technological imbalances among their carriers. However, now this trend also creates risks for all users of large language models (Large Language Model, LLM), since publicly -accessible API machine translations can bypass LLM security measures.
Openai recognized the significance of the study and announced the intention to take into account its results. This emphasizes the need to search for complex solutions to ensure the safety of AI, including improving teaching models in limited resources and developing more effective filtration mechanisms.