为应对人工智能的快速发展，顶尖的 AI 主管与科学家，包括 Google DeepMind 的 Demis Hassabis、OpenAI 的 Sam Altman、Anthropic 的 Dario Amodei 以及 Microsoft AI 的 Mustafa Suleyman，共同签署了一封公开信，敦促美国立法者针对合成 DNA 与 RNA 的供应商制定强制性的筛查法律。自从 Arthur Kornberg 于 1950 年代（1950s）首次成功合成 DNA 以来，该过程已在全球范围内实现自动化与高度商业化。目前，尽管许多供应商会对客户和订单进行筛查，但仍缺乏全球统一的强制规范，这引发了人们对于不良意图者可能利用 AI 工具规避自愿性检查并获取危险基因物质的担忧。
未受监管的基因合成所带来的潜在风险在 2017 年（2017）得到了关注，当时加拿大研究人员利用价值 100,000 美元（100,000 USD）的邮购 DNA 成功重建了已灭绝的马痘病毒，引发了人们对天花等致命病毒可能被重新制造的担忧。尽管生物恐怖袭击在历史上非常罕见，但大型语言模型与 AI 蛋白质设计工具的融合，现在使用户能够快速生成危险的毒素和病原体。至关重要的是，微软研究人员发表的一项研究表明，AI 工具能够通过模拟安全蛋白质的结构，生成可完全绕过现有筛查软件的潜在危险基因序列。
为了应对这些漏洞，美国参议院于今年早些时候提出了一项两党法案，要求所有美国境内的基因合成供应商必须对客户进行背景审查并筛查订单，这扩展了先前仅适用于接受联邦资助机构的联邦指南。然而，David Relman 和 Geoff Ralston 等安全专家强调，仅靠硬件层面的 DNA 筛查是不够的。由于筛查系统在某些情况下可能会失效，AI 公司必须承担起责任，在模型内部实施控制措施，以阻止用户生成有关危险生物制剂的指令或设计。
In response to the rapid advancement of artificial intelligence, leading AI executives and scientists, including Google DeepMind’s Demis Hassabis, OpenAI’s Sam Altman, Anthropic’s Dario Amodei, and Microsoft AI’s Mustafa Suleyman, signed a public letter urging US lawmakers to establish mandatory screening laws for synthetic DNA and RNA providers. Since Arthur Kornberg first synthesized DNA in the 1950s, the process has become automated and highly commercialized globally. Currently, while many providers screen customers and orders, there is no universal mandate, raising fears that bad actors could exploit AI tools to bypass voluntary checks and obtain dangerous genetic materials.
The potential risks of unregulated gene synthesis were highlighted in 2017 when Canadian researchers successfully reconstituted the extinct horsepox virus using mail-order DNA costing $100,000 (approximately 100,000 USD), sparking concerns that a deadly pathogen like smallpox could be recreated. Although bioterror attacks remain historically rare, the integration of large language models and AI protein design tools now allows users to quickly generate dangerous toxins and pathogens. Crucially, a study by Microsoft researchers demonstrated that AI tools could design potentially dangerous gene sequences that completely evaded existing screening software by mimicking safe protein structures.
To address these vulnerabilities, the US Senate introduced a bipartisan bill earlier this year requiring all domestic gene synthesis providers to vet customers and screen orders, expanding upon previous federal guidelines that only applied to entities receiving federal funding. However, security experts like David Relman and Geoff Ralston emphasize that hardware-level DNA screening is insufficient. Because screening systems can fail, AI companies must step up by implementing internal controls within their models to prevent users from generating instructions or designs for dangerous biological agents.