English (United Kingdom)

https://curated-unify.zendy.io/wp-json/zendy-region/v1/featured_content/oa?rat=en

https://curated-unify.zendy.io/wp-json/zendy-region/v1/highlighted_journal/

Global: Zendy Tools annual

Presents the keyphrase highlighting as premium feature

Keyphrase Highlighting

Presents the summarisation as premium feature

Summarisation

Insights

Presents the pdf analysis as premium feature

PDF Analysis

Presents the zaia usage as premium feature

ZAIA

Global: Zendy Tools monthly

Global: Zendy Plus monthly

Presents the access of premium content as premium feature

Premium Content

Global: Zendy Plus annual

Global: Zendy Free Plan

While large language models (LLMs) exhibit remarkable capabilities across awide range of tasks, they pose potential safety concerns, such as the``jailbreak'' problem, wherein malicious instructions can manipulate LLMs toexhibit undesirable behavior. Although several preventive measures have beendeveloped to mitigate the potential risks associated with LLMs, they haveprimarily focused on English data. In this study, we reveal the presence ofmultilingual jailbreak challenges within LLMs and consider two potential riskscenarios: unintentional and intentional. The unintentional scenario involvesusers querying LLMs using non-English prompts and inadvertently bypassing thesafety mechanisms, while the intentional scenario concerns malicious userscombining malicious instructions with multilingual prompts to deliberatelyattack LLMs. The experimental results reveal that in the unintentionalscenario, the rate of unsafe content increases as the availability of languagesdecreases. Specifically, low-resource languages exhibit three times thelikelihood of encountering harmful content compared to high-resource languages,with both ChatGPT and GPT-4. In the intentional scenario, multilingual promptscan exacerbate the negative impact of malicious instructions, withastonishingly high rates of unsafe output: 80.92\% for ChatGPT and 40.71\% forGPT-4. To handle such a challenge in the multilingual context, we propose anovel \textsc{Self-Defense} framework that automatically generates multilingualtraining data for safety fine-tuning. Experimental results show that ChatGPTfine-tuned with such data can achieve a substantial reduction in unsafe contentgeneration. Data is available athttps://github.com/DAMO-NLP-SG/multilingual-safety-for-LLMs. Warning: Thispaper contains examples with potentially harmful content.

Multilingual Jailbreak Challenges in Large Language Models