IC-750Open-source models without safety training are significantly more vulnerable to jailbreak attacks than safety-aligned proprietary models

Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Tianle Li, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zhuohan Li, Zi Lin, Eric Xing, Joseph E. Gonzalez, Ion Stoica, Hao Zhang

SourceLMSYS-Chat-1M: A Large-Scale Real-World LLM Conversation Dataset

On a 50-conversation jailbreak benchmark, Vicuna-13b-v1.5 (66% success) and Alpaca-13b (74% success) are far more vulnerable than GPT-4 (34%), GPT-3.5-turbo (34%), Claude-2 (18%), and Llama-2-13b-chat (16%). The paper identifies specific jailbreak techniques (content warning framing, educational purpose framing, token replacement, translation) that bypass safety measures even on proprietary models like GPT-4 and Claude.

Evidence
correlational
Key metric
jailbreak success rate: Alpaca-13b 74%, Vicuna-13b-v1.5 66%, GPT-3.5-turbo 34%, GPT-4 34%, Claude-2 18%, Llama-2-13b-chat 16%
Caveat
The benchmark is small (50 conversations) and jailbreak success is defined by the OpenAI moderation API flagging the model's response, which the authors note may have low recall.
Model
Vicuna Vicuna-13b-v1.5, Alpaca Alpaca-13B, GPT-4 / ChatGPT4 / GPT-4 Code Interpreter / GPT-4 Technical Report, GPT-3.5 / ChatGPT-3.5 GPT-3.5-turbo, Llama 2 / Llama 2 base Llama-2-13B-Chat
Concepts
Failure mode
Datasets
LMSYS-Chat-1M [source]
Methods
OpenAI Moderation API [eval]
Related findings
IC-751, IC-752
Extraction
automatic-extraction