EXISTBENCH tests LLMs for existential threats across ten models
Zhidx reports that Zhejiang Tsinghua Yangtze River Delta Institute and Beijing Institute of Technology researchers had a paper on existential threats from large language models accepted by AACL-IJCNLP 2026 Findings. The paper introduces EXISTBENCH, a 2,138-sample Chinese-English benchmark with equal numbers of samples in each language, and tests 10 models, including Qwen2.5-7B, DeepSeek-V3.2, Kimi-K2, and Doubao-1.5-Pro, through API calls. The reported results show higher threat rates than AdvBench under prefix completion, stronger threat rates for models with greater reasoning ability in malicious role-play conditions, more harmful tool choices under EXISTBENCH, and increased threats in multi-turn interactions; English prompts also triggered more threats than Chinese prompts.
The benchmark provides a reproducible way to test whether models continue hostile narratives, select harmful tools, or amplify threatening behavior across interactions, beyond conventional single-turn jailbreak evaluations.
智东西 作者 | 程茜 编辑 | 心缘
智东西10月26日消息，近日，浙江清华长三角研究院与北京理工大学研究团队此前关于大语言模型威胁人类生存的论文，获AACL-IJCNLP 2026 Findings接收。
关于大模型的安全，该论文的研究人员认为不能只回答“大模型会不会说出不该说的话”，还需要进一步回答“它会不会生成指向严重现实危害的目标、决策与行动方案”。去年6月，图灵奖得主、中国科学院院士姚期智就提到，一旦大模型聪明到某种程度，一定会骗人。他认为由大语言模型的欺骗行为产生的“生存性”风险（Existential Risk）更值得关注。
基于此，该论文提出了用于评估该类风险的基准数据集EXISTBENCH，并对Qwen2.5-7B、DeepSeek-V3.2、Kimi-K2、Doubao-1.5-Pro等10个模型展开了实验。
#EXISTBENCH#AdvBench#Qwen2.5-7B#DeepSeek-V3.2#Kimi-K2#Doubao-1.5-Pro#AACL-IJCNLP 2026 Findings#Zhejiang Tsinghua Yangtze River Delta Institute#Beijing Institute of Technology