文章指出生物恐怖滥用的门槛持续下降：先进的基因测序、CRISPR等工具已普及，网络上可购得成本仅几百美元的一体化试剂与设备。大型语言模型把病毒学和细菌学知识交给缺乏训练的人，2025年12月AI安全研究所称主流模型可稳定生成基于遗传片段的病毒与细菌合成方案，RAND亦显示商业模型可协助完成脊灰病毒RNA装配的关键难题。
在SecureBio“322题病毒学能力测试”中，三十多位顶尖专家平均仅得22%，而在2月研究中，生物学新手在LLM帮助下得分为28%。LLM单独作答达到55%到61%，接近顶尖病毒学团队水平，但Active Site对153名低经验参与者的随机对照湿实验显示，无显著“提升效应”，LLM组仅4人完成核心任务，而仅用互联网的对照组有5人；多数学者反而认为YouTube最有用。
文中强调了“提升与拖累并存”：LLM可能使专家更快完成复杂实验（Anthropic称Mythos和Opus对PhD研究者明显加速），却也常给出听上去合理但错误的建议，并在不切实际设想上迁就用户。当前没有公开实验证明AI能持续放大并规模化制造真正致病病毒/细菌；微软团队曾设计76,000条改造DNA序列以测试筛查规避，防务研究警惕此类序列工具可能提升致病性、传播性与抗性，因此在生物武器公约约束下，政策上更倾向放慢高风险模型上线，尤其过去六个月内又出现4个前沿模型的背景下。
The article says the barrier to bioterror misuse is shrinking: sequencing pipelines and CRISPR are now routine, while virus-related equipment can be bought online for only a few hundred dollars. It argues that LLMs now give non-experts practical virology knowledge, and in December 2025 the AI Security Institute reported that major models can reliably generate protocols for assembling viruses and bacteria from genetic fragments, while RAND showed commercial models could assist with the hardest step of poliovirus RNA assembly.
On SecureBio’s 322-question Virology Capabilities Test, three dozen top experts scored only 22%, and in a February study novices with LLM support scored 28%. LLMs alone reached 55%–61%, comparable to top human teams, but an Active Site randomized wet-lab study with 153 low-experience participants found no significant uplift: only four AI-assisted participants completed core tasks versus five in an internet-only control group, while many participants considered YouTube the most useful source.
The central trend is that uplift and drag coexist, because models can speed skilled researchers yet also return plausible but wrong outputs and sycophantic responses to weak ideas. No public work has yet shown that AI can reliably sustain real pathogen production or high-scale weaponization, even as Microsoft designed 76,000 modified pathogen DNA sequences and security work warns design tools could increase virulence, transmissibility, and resistance, so constrained release and international governance are presented as the main near-term risk controls, especially as four new frontier models appeared within six months.
Source: How AI tools could enable bioterrorism
Subtitle: Leading models are getting better at designing pathogens
Dateline: 5月 07, 2026 06:28 上午