Scientists hacked GPT-4 using rare languages: Zulu, Gaelic, Hmong and Guarani

[or1k]

The experts believe (https://www.theregister.com/2024/01/31/gpt4_gaelic_safety/?td=rt-4a) that they managed to bypass the GPT-4 security filters that prevent them from delivering "unsafe content". They translated the queries into rare languages and translated the answers back into English, and they looked quite convincing.

✍️ The scientists ran 520 unsafe requests through GPT-4 and found that they were able to bypass the protection 79% of the time. Simple translation was almost as successful at "cracking" the large language model as other, more sophisticated and technologically advanced methods. The model was more likely to fulfill queries related to terrorism, financial crimes, and misinformation.

By the way, resource-limited languages are now spoken by about 1.2 billion people.