Diversity Helps Jailbreak Large Language Models
Fuente:
arXiv
Gespeichert in:
| Hauptverfasser: | , , , , |
|---|---|
| Format: | Preprint |
| Veröffentlicht: |
2024
|
| Schlagworte: | |
| Online-Zugang: | |
| Tags: |
Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
|
| _version_ | 1866913830722863104 |
|---|---|
| author | Zhao, Weiliang Ben-Levi, Daniel Hao, Wei Yang, Junfeng Mao, Chengzhi |
| author_facet | Zhao, Weiliang Ben-Levi, Daniel Hao, Wei Yang, Junfeng Mao, Chengzhi |
| contents | We have uncovered a powerful jailbreak technique that leverages large language models' ability to diverge from prior context, enabling them to bypass safety constraints and generate harmful outputs. By simply instructing the LLM to deviate and obfuscate previous attacks, our method dramatically outperforms existing approaches, achieving up to a 62.83% higher success rate in compromising ten leading chatbots, including GPT-4, Gemini, and Llama, while using only 12.9% of the queries. This revelation exposes a critical flaw in current LLM safety training, suggesting that existing methods may merely mask vulnerabilities rather than eliminate them. Our findings sound an urgent alarm for the need to revolutionize testing methodologies to ensure robust and reliable LLM security. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2411_04223 |
| institution | arXiv |
| publishDate | 2024 |
| record_format | arxiv |
| spellingShingle | Diversity Helps Jailbreak Large Language Models Zhao, Weiliang Ben-Levi, Daniel Hao, Wei Yang, Junfeng Mao, Chengzhi Computation and Language We have uncovered a powerful jailbreak technique that leverages large language models' ability to diverge from prior context, enabling them to bypass safety constraints and generate harmful outputs. By simply instructing the LLM to deviate and obfuscate previous attacks, our method dramatically outperforms existing approaches, achieving up to a 62.83% higher success rate in compromising ten leading chatbots, including GPT-4, Gemini, and Llama, while using only 12.9% of the queries. This revelation exposes a critical flaw in current LLM safety training, suggesting that existing methods may merely mask vulnerabilities rather than eliminate them. Our findings sound an urgent alarm for the need to revolutionize testing methodologies to ensure robust and reliable LLM security. |
| title | Diversity Helps Jailbreak Large Language Models |
| topic | Computation and Language |
| url | https://arxiv.org/abs/2411.04223 |