What AI can and cannot do in biosecurity

Things have gotten intense with AI in the past week.

过去一周，人工智能的发展变得异常激烈。

People working in AI, from researchers to CEOs, think it could threaten humanity.

从研究人员到首席执行官，从事人工智能工作的人们都认为它可能威胁人类。

One researcher put the probability of doom, as those in the field call it, at over 10%: the chance that mankind is wiped out by 2030.

一位研究人员认为，业内人士所说的“厄运”概率超过 10%：人类到 2030 年将被消灭。

Is this accurate?

这准确吗？

Is this a ploy to hype AI powers ahead of an IPO valuation?

这是在 IPO 估值之前炒作人工智能力量的策略吗？

Or maybe a distraction from other unpopular AI issues, such as amassing power and wealth and disproportionately using natural resources?

或者可能会分散人们对其他不受欢迎的人工智能问题的注意力，例如积累权力和财富以及过度使用自然资源？

I don’t know; it’s probably somewhere in the middle of it all.

我不知道;它可能就在这一切中间的某个地方。

Underneath the recent debate over hypothetical scenarios lies a fascinating, critical public health question: Where does biosecurity risk actually intersect with superintelligence?

最近关于假设情景的争论背后隐藏着一个令人着迷的关键公共卫生问题：生物安全风险实际上与超级智能有何交叉？

I texted my friend Dr. Claire Dillavou out of pure curiosity.

出于纯粹的好奇，我给我的朋友克莱尔·迪拉沃博士发了短信。

She is a highly trained applied epidemiologist and has been in the tech and AI world for a while now.

她是一位训练有素的应用流行病学家，已经在科技和人工智能领域工作了一段时间。

We got to talking through feasibility, challenges, and the solutions needed from a health perspective.

我们从健康角度讨论了可行性、挑战和所需的解决方案。

We thought we’d bring the YLE community along for the ride, too.

我们认为我们也应该让 YLE 社区也参与进来。

Here’s what we know, what we don’t know, and what the field is doing to try to close that gap.

以下是我们所知道的、我们不知道的以及该领域正在采取的措施来缩小这一差距。

Biosecurity is a massive field.

生物安全是一个广阔的领域。

Threats can range from something natural, like a virus or an antibiotic-resistant bug, to something engineered, like a toxin modified to be more potent.

威胁的范围可以从天然的东西（如病毒或抗生素耐药性细菌）到工程的东西（如经过修改以增强效力的毒素）。

All of these have different probabilities and feasibility of actually happening, with or without AI.

无论有或没有人工智能，所有这些实际发生的概率和可行性都不同。

Biological trade-offs are real.

生物学上的权衡是真实存在的。

Changing nature can be very hard.

改变自然可能非常困难。

Take viruses.

以病毒为例。

They face a fundamental biological trade-off: highly lethal or highly contagious, but typically not both.

它们面临着基本的生物学权衡：高度致命或高度传染性，但通常不会两者兼而有之。

Ebola, for example, can have a fatality rate of up to 90% , but it only spreads through close contact with other people.

例如，埃博拉病毒的死亡率高达 90%，但它只能通过与其他人的密切接触传播。

When a virus kills its host (like a human), it can’t spread widely to other people.

当病毒杀死宿主（如人类）时，它就无法广泛传播给其他人。

Measles, by contrast, is the most contagious virus on earth.

相比之下，麻疹是地球上传染性最强的病毒。

It’s a horrible disease, but it doesn’t have a high mortality rate.

这是一种可怕的疾病，但死亡率并不高。

This tradeoff isn’t impossible to get around, but it is very hard to go against biology.

这种权衡并非不可能避免，但很难违背生物学原理。

AI can lower barriers to knowledge.

人工智能可以降低知识壁垒。

Tasks that used to require years of specialized training, a well-funded lab, or slow trial-and-error can now be meaningfully accelerated by current AI systems.

过去需要多年专业培训、资金充足的实验室或缓慢试错的任务现在可以通过当前的人工智能系统得到有意义的加速。

This includes tasks such as literature synthesis (ask any graduate student how painful this was!), predicting how proteins fold, designing novel and efficient gene-editing tools, and helping design and facilitate experiments.

这包括诸如文献综合（问问任何研究生这有多痛苦！）、预测蛋白质如何折叠、设计新颖且高效的基因编辑工具以及帮助设计和促进实验等任务。

While these are amazing advances, they are built on a compendium of physical work in a lab, actually growing, modifying, and handling dangerous pathogens.

虽然这些都是惊人的进步，但它们是建立在实验室的体力工作概要之上的，实际上是培养、修改和处理危险的病原体。

This physical work will remain necessary for the foreseeable future.

在可预见的未来，这项体力劳动仍然是必要的。

Unsurprisingly, some people are already trying to use AI to do harm .

毫不奇怪，有些人已经开始尝试利用人工智能来造成伤害。

Anthropic recently published five case studies exploring the use of Claude to assist with biological weapons development.

Anthropic 最近发表了五个案例研究，探索利用克劳德协助生物武器的开发。

Several centered on that same lethality/contagiousness trade-off:

一些集中在相同的杀伤力/传染性权衡上：

Making a virus spread more easily between people and dodge the immune system (chikungunya virus).

使病毒更容易在人与人之间传播并躲避免疫系统（基孔肯雅病毒）。

This work may have been tied to a state or military research group.

这项工作可能与国家或军事研究小组有关。

Helping a virus jump to a new species (bird flu).

帮助病毒跳跃到新物种（禽流感）。

Someone used Claude to help bird flu adapt so it could spread more easily between mammals (a step toward spreading in humans).

有人利用克劳德来帮助禽流感适应，这样它就可以更容易地在哺乳动物之间传播（这是向人类传播迈出的一步）。

Making toxins more dangerous (venom peptides).

使毒素更加危险（毒液肽）。

Someone used Claude to build a big catalog of venom toxins and a system to make them more harmful.

有人利用克劳德建立了一个毒液毒素的大目录和一个使它们更具危害性的系统。

Studying how viruses dodge immune defenses (orthopoxviruses) like studying the genes that control how the immune system fights off viruses, like smallpox.

研究病毒如何躲避免疫防御（正痘病毒），就像研究控制免疫系统如何抵抗天花等病毒的基因一样。

Redesigning toxins to make them more dangerous, like a virus that causes severe, deadly bleeding.

重新设计毒素，使其更加危险，就像导致严重、致命出血的病毒一样。

Anthropic detected this activity, banned the accounts involved, and used what it learned to strengthen safeguards in newer models.

Anthropic 发现了这一活动，禁止了相关账户，并利用所学到的知识来加强新模型的安全保障。

But all of this was sped-up knowledge, and it doesn’t get around the needs of the physical world for this to actually be executed.

但所有这些都是加速的知识，它并没有绕过物理世界实际执行的需求。

But the same capabilities cut the other way .

但同样的能力也有相反的效果。

AI can also help us:

人工智能还可以帮助我们：

Decipher the makeup of a novel virus and help us develop a vaccine or treatment more quickly and then mass-produce that solution faster.

破译新型病毒的构成，帮助我们更快地开发疫苗或治疗方法，然后更快地批量生产该解决方案。

Help clinicians diagnose diseases with higher confidence and more rapidly.

帮助临床医生更有信心、更快速地诊断疾病。

We’ve seen this in a number of recent articles.

我们在最近的许多文章中都看到了这一点。

Predict individual disease risk years in advance before symptom onset to enhance prevention and slow or stop disease progression.

在症状出现前数年预测个体疾病风险，以加强预防并减缓或阻止疾病进展。

Provide higher-quality, more affordable, and easily accessible health care for the entire population.

为全体人民提供更高质量、更负担得起且更容易获得的医疗保健。

The leap from “AI assists with biology” to “AI enables pandemic-scale bioweapons” involves many steps, and experts have big, unanswered questions:

从“人工智能协助生物学”到“人工智能实现大流行规模的生物武器”的飞跃涉及许多步骤，专家们有一些尚未解答的重大问题：

How much of the barrier to creating dangerous pathogens is knowledge versus wet lab skill, equipment access, and iteration time?

创造危险病原体的障碍有多少是知识与湿实验室技能、设备访问和迭代时间？

AI primarily helps with the former.

人工智能主要帮助解决前者。

Will truly novel engineered pathogens be more or less dangerous than naturally occurring ones?

真正的新型工程病原体比自然存在的病原体危险程度更高还是更低？

How effective are the government's and the Frontier model companies’ existing biosurveillance and biosecurity infrastructure as countermeasures?

政府和前沿模范公司现有的生物监测和生物安全基础设施作为对策的效果如何？

There is an active debate in the expert community.

专家界对此展开了激烈的争论。

Some biosecurity researchers treat AI-bio risk as a near-term serious concern.

一些生物安全研究人员将人工智能生物风险视为近期的严重问题。

Others argue that the physical and institutional barriers remain the dominant constraint.

其他人则认为，物质和制度障碍仍然是主要制约因素。

We can’t honestly say there is a consensus.

我们不能诚实地说存在共识。

AI plausibly increases the risk at the margin.

人工智能似乎会增加边际风险。

Whether that increase is small or catastrophic depends on factors such as AI capability trajectories, biosecurity responses, governance, etc.

这种增长是小还是灾难性的，取决于人工智能能力轨迹、生物安全响应、治理等因素。

And all of this ultimately leads to the large, unanswered question: Can/will the benefits of AI outweigh the risks in biosecurity?

所有这一切最终导致了一个悬而未决的大问题：人工智能的好处能否/将会超过生物安全的风险？

Nobody has a clear answer yet to these questions, but several efforts are underway to close the gap between “we’re worried” and “we actually know.” Much work needs to be done, but these three priorities come to mind as some of the most urgent in the biosecurity world:

目前还没有人对这些问题有明确的答案，但我们正在采取一些努力来缩小“我们担心”和“我们实际上知道”之间的差距。需要做很多工作，但以下三个优先事项是生物安全领域最紧迫的优先事项：

Protect data.

保护数据。

We know that the most recent Frontier models are so advanced they can easily hack into weak systems, access highly protected data, cover their tracks, and create entire systems on their own to protect themselves, as seen in the recent OpenAI Hugging Face incident, among others.

我们知道，最新的 Frontier 模型非常先进，它们可以轻松侵入薄弱的系统、访问高度保护的数据、掩盖自己的踪迹，并自行创建整个系统来保护自己，正如最近的 OpenAI Hugging Face 事件等。

How can we really protect personal and sensitive data in a meaningful way?

我们如何才能真正以有意义的方式保护个人和敏感数据？

And who owns or has a right to use these data, even if de-identified?

即使这些数据已被去识别化，谁拥有或有权使用这些数据？

Pressure for more transparency.

要求提高透明度的压力。

The reality of the gap between what the general public knows/has seen and what’s happening in restricted areas of these companies is significant and widening daily.

公众所知道/所看到的情况与这些公司限制领域内发生的情况之间的现实差距是巨大的，而且每天都在扩大。

It is hard for the public and legislators to advocate for specific policies without more transparency and insight.

如果没有更多的透明度和洞察力，公众和立法者就很难倡导具体政策。

De-escalate the geopolitical race.

缓和地缘政治竞赛。

The geopolitical tension at the center of the rapid AI progression is a genuine unresolved problem that needs more thought leadership, diplomacy, and action if true “pacing” of this technology is to take place.

人工智能快速发展的核心地缘政治紧张局势是一个真正尚未解决的问题，如果要实现这项技术的真正“节奏”，就需要更多的思想领导、外交和行动。

AI is fundamentally changing everything from how we do good and important public health and health care work to how we develop biosurveillance systems that keep pace with the rapid escalation of capabilities enabled by this technology.

人工智能正在从根本上改变一切，从我们如何做好重要的公共卫生和医疗保健工作，到我们如何开发生物监测系统，以跟上该技术所实现的能力的快速升级。

There will always be bad actors trying to abuse novel capabilities enabled by AI, but at this moment we are still in control of what we allow to happen.

总会有坏人试图滥用人工智能带来的新功能，但目前我们仍然可以控制我们允许发生的事情。

This window is closing rapidly, so we need to absolutely double down on security, safety, and legislative efforts while we reinforce our biosurveillance systems and ensure our biosecurity measures are well-articulated, practiced, and funded.

这个窗口正在迅速关闭，因此我们绝对需要在安全和立法方面加倍努力，同时加强我们的生物监测系统并确保我们的生物安全措施得到充分阐述、实施和资助。

Claire Dillavou, PhD MPH, is a trained infectious disease and behavioral epidemiologist who has worked in applied public health at every level of domestic government, with international governments, and in the private sector in startups and consulting.

Claire Dillavou，公共卫生硕士，是一位训练有素的传染病和行为流行病学家，曾在国内各级政府、国际政府以及私营部门的初创企业和咨询机构从事应用公共卫生工作。

Your Local Epidemiologist (YLE) comprises a team of experts, ranging from physicians to immunologists to epidemiologists to nutritionists, working together with one goal: to “translate” ever-evolving public health science so that people are well-equipped to make evidence-based decisions.

您的当地流行病学家 (YLE) 由一个专家团队组成，其中包括医生、免疫学家、流行病学家和营养学家，他们共同努力实现一个目标：“转化”不断发展的公共卫生科学，以便人们有能力做出基于证据的决策。

The YLE suite of newsletters reaches over 475,000 people across more than 132 countries.

YLE 新闻通讯套件覆盖超过 132 个国家/地区的 475,000 多人。

This newsletter is free to everyone, thanks to the generous support of fellow YLE community members.

感谢 YLE 社区成员的慷慨支持，本通讯免费向所有人开放。

To support the effort, subscribe or upgrade below:

为了支持这项工作，请订阅或升级如下：

Appreciate this overview.

欣赏这个概述。

Just a note on linking to METR's report, though: there are concerns about significant financial and personal conflicts of interest between them and the AI companies.

不过，请注意与 METR 报告的链接：人们担心他们与人工智能公司之间存在重大财务和个人利益冲突。

"Straight out of the tobacco & fossil fuel playbook.

“直接出自烟草和化石燃料手册。

Have the same companies and billionaires fund "independent" institutes staffed by the same people, and "cite" their reports as "independent experts" showing "consensus.""

让相同的公司和亿万富翁资助由相同人员组成的“独立”机构，并“引用”他们的报告作为“独立专家”，显示“共识”。

"it's all the same EA people with the same open philanthropy funding.

“都是同样的 EA 人员，拥有同样的公开慈善资金。

they are not "independent" by any definition.

从任何定义来看，它们都不是“独立的”。

- One of the most prominent researchers is married to someone on the board of OAI."

- 一位最著名的研究人员与 OAI 董事会成员结婚。”

"I applaud this but any mature audit ecosystem has *auditor conduct standards*.

“我对此表示赞赏，但任何成熟的审计生态系统都有*审计员行为标准*。

Was surprised to learn a METR investigator for the HF/OpenAI incident is married to an OpenAI board member, many access contracts are highly constrained etc. PCAOB in accounting exists for good reason!"

惊讶地发现 HF/OpenAI 事件的 METR 调查员与 OpenAI 董事会成员结婚，许多访问合同受到高度限制等。会计领域的 PCAOB 存在是有充分理由的！”

It's all about political will, and constraining the bad impulses and bad decisions of humans chasing their own narrow interests.

这一切都与政治意愿有关，并限制人类追逐自身狭隘利益的不良冲动和错误决定。

Thanks for this well written explainer.

感谢这位写得好的解释者。