arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2601.12652cs.CY
收藏

大语言模型部署中的伦理风险:医疗伦理“ Jailbreak”评估

Ethical Risks in Deploying Large Language Models: An Evaluation of Medical Ethics Jailbreaking

Shanghai Jiaotong University(上海交通大学);Fudan University(复旦大学)

Chutian Huang, Dake Cao, Jiacheng Ji, Yunlou Fan, Chengze Yan, Hanhui Xu

AI总结:

本文评估了大语言模型在医疗伦理领域面临的安全风险,发现七种主流模型在对抗测试中表现不一,其中Claude-Sonnet-4-Reasoning表现最稳健,而其他五种模型几乎全部失效。

摘要:

背景:尽管大语言模型(LLMs)已广泛应用,但恶意提示工程特别是“jailbreak攻击”通过诱导模型绕过内部安全机制,带来了严重安全风险。当前的基准测试主要关注公共安全和西方文化规范,留下了一个关键空白,即在中文语境下评估医疗伦理这一小众、高风险领域的缺口。目标:建立专门的jailbreak评估框架,用于评估中国医疗伦理的防御韧性和伦理一致性,系统评估七种主要的LLMs在面对复杂对抗模拟时的防御性能和伦理对齐性。方法:我们使用DeepInception框架中的“角色扮演+场景模拟+多轮对话”向量,评估了七种主要模型(例如GPT-5、Claude-Sonnet-4-Reasoning、DeepSeek-R1)。测试聚焦于八个高风险主题,包括商业代孕和器官交易,利用分层评分矩阵量化攻击成功率(ASR)和ASR增益。结果:观察到防御系统全面崩溃,尽管模型基线合规性较高,但jailbreak ASR达到82.1%,代表ASR增益超过80个百分点。Claude-Sonnet-4-Reasoning成为最稳健的模型,而包括Gemini-2.5-Pro和GPT-4.1在内的五种模型几乎全部失败,ASR在96%至100%之间。结论:当前LLMs在医疗伦理方面高度易受上下文操纵,通常优先考虑“有用性”而非安全约束。为了提高安全性,我们建议从结果监督转向过程监督,实施多因素身份验证,并建立跨模型“联合防御”机制。

英文摘要:

Background: While Large Language Models (LLMs) have achieved widespread adoption, malicious prompt engineering specifically "jailbreak attacks" poses severe security risks by inducing models to bypass internal safety mechanisms. Current benchmarks predominantly focus on public safety and Western cultural norms, leaving a critical gap in evaluating the niche, high-risk domain of medical ethics within the Chinese context. Objective: To establish a specialized jailbreak evaluation framework for Chinese medical ethics and to systematically assess the defensive resilience and ethical alignment of seven prominent LLMs when subjected to sophisticated adversarial simulations. Methodology: We evaluated seven prominent models (e.g., GPT-5, Claude-Sonnet-4-Reasoning, DeepSeek-R1) using a "role-playing + scenario simulation + multi-turn dialogue" vector within the DeepInception framework. The testing focused on eight high-risk themes, including commercial surrogacy and organ trading, utilizing a hierarchical scoring matrix to quantify the Attack Success Rate (ASR) and ASR Gain. Results: A systemic collapse of defenses was observed, whereas models demonstrated high baseline compliance, the jailbreak ASR reached 82.1%, representing an ASR Gain of over 80 percentage points. Claude-Sonnet-4-Reasoning emerged as the most robust model, while five models including Gemini-2.5-Pro and GPT-4.1 exhibited near-total failure with ASRs between 96% and 100%. Conclusions: Current LLMs are highly vulnerable to contextual manipulation in medical ethics, often prioritizing "helpfulness" over safety constraints. To enhance security, we recommend a transition from outcome to process supervision, the implementation of multi-factor identity verification, and the establishment of cross-model "joint defense" mechanisms.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑