Researchers Find Gaps in GPT-4 Security by Translating English Prompts to Uncommon Languages

Researchers at Brown University found vulnerabilities in GPT-4's security by translating unsafe English prompts into uncommon languages like Zulu and Gaelic.

Translated prompts allowed them to bypass restrictions 79% of the time versus 1% in English.

Prompts translated to Zulu asking how to shoplift evaded GPT-4's usual "I can't assist with that" response in English.

Researchers emphasized the need to include diverse languages beyond just English in future AI safety testing.

Before publishing, researchers shared findings with OpenAI to mitigate risks of giving ideas to bad actors.