字号·· | 护眼
彭博社

美国如何挺过人工智能危险区How the US Can Survive the AI Danger Zone

点「原文对照」整页切到原文,或双击某段只看那段的原文。

危险的十字路口。Dangerous intersection.

这原本是正义一方的胜利。10月,两家网络安全公司披露,它们利用先进的人工智能模型发现了PHP编程语言中的关键缺陷,该语言支撑着互联网70%的基础设施。随后,它们在数周内悄无声息地协助开发修复补丁,然后才公开披露。

It started as a win for the good guys. In October, two cybersecurity firms disclosed that they had used advanced AI models to uncover critical flaws in PHP, a programming language that underpins 70% of the internet. They had then worked quietly for weeks to help develop fixes before going public.

然而,这一公开确认——因不合时宜的泄露而被迫提前——引发了一场绝望的竞赛,各国政府和企业争相修补漏洞,以防不法分子加以利用。日本首相高市早苗敦促企业不惜一切代价——必要时甚至暂停服务——以限制其风险敞口。

Yet that public acknowledgment — accelerated by untimely leaks — touched off a desperate race as governments and corporations scrambled to patch flaws before nefarious actors could exploit them. Japan’s prime minister, Sanae Takaichi, urged companies to pull out all the stops — suspending service if necessary — to limit their exposure.

这一事件凸显了一个可能性:罪犯、恐怖分子或流氓国家可能利用人工智能模型,借助毁灭性的网络漏洞发动攻击。它也凸显了人工智能革命中一个极其危险的时刻已经到来:与人工智能相关的冲击显现速度,远快于负责管理这些冲击者的应对速度。

The episode raised the possibility that criminals, terrorists or rogue states might use AI models to take advantage of crippling cyber-vulnerabilities. It also underscored that a deeply perilous moment in the artificial intelligence revolution has arrived: AI-related disruptions are manifesting far faster than those charged with managing them can respond.

网络领域就是这一问题的典型例证。新的人工智能模型擅长发掘代码中长期隐藏、可能造成灾难性后果的弱点。Anthropic公司的Mythos模型发现了一个经受住27年人工审查的漏洞。Anthropic的一项网络安全计划“Project Glasswing”在4月至7月期间发现了12.9万个经验证的漏洞。其中3.3万多个被认定为“关键”或“高危”级别。

The cyber domain exemplifies the problem. New AI models excel at discovering long-buried and potentially catastrophic weaknesses in code. Anthropic’s Mythos found a bug that had withstood 27 years of human review. An Anthropic cybersecurity initiative, Project Glasswing, turned up 129,000 verified vulnerabilities between April and July. More than 33,000 of those flaws were deemed “critical” or “high-severity.”

从技术本质上讲,没有理由认为人工智能对网络攻击者比对网络防御者更有利:Glasswing旨在将尖端能力优先提供给受信任的组织。挑战在于,防御者往往受制于过时的政策、资源不足,以及关键漏洞被发现到修复之间的时间差。

There’s no inherent technological reason that AI should be more advantageous for cyberattackers than for cyberdefenders: Glasswing aims to get cutting-edge capabilities to the trusted organizations first. The challenge is that defenders are often constrained by out-of-date policies, insufficient resources and the lag between discovery of a critical vulnerability and its repair.

这些限制助长了敏捷的对手——例如寻求赎金的罪犯,或试图在冲突中瘫痪美国社会的敌对国家——将 AI 武器化,通过自动化和加速利用不安全网络所需的步骤。“进攻能力正在到来,”前国防部网络政策助理部长迈克尔·萨尔迈耶在《》(一本我于今年早些时候编辑的书)中写道。“问题是防御能否跟上步伐。”同样的问题也适用于 AI 赋能的生物威胁。

Those constraints help agile adversaries —crooks seeking ransoms, say, or hostile states trying to paralyze American society during a conflict — weaponize AI by automating and accelerating steps needed to exploit insecure networks. “The offensive capabilities are arriving now,” wrote Michael Sulmeyer, former assistant secretary of defense for cyber policy, in , a book I edited earlier this year. “The question is whether the defense will keep pace.”The same question pertains to AI-enabled biothreats.

拜登政府的国土安全顾问伊丽莎白·舍伍德-兰德尔警告了一个令人毛骨悚然的场景:一个恐怖组织使用禁用了安全控制的 AI 模型设计出一种致命、高传染性的禽流感毒株。随后,该组织利用合同服务实验室生产病毒并将其释放,对美国人口和粮食供应造成毁灭性打击。

Elizabeth Sherwood-Randall, the Joe Biden administration’s homeland security adviser, has warned of a chilling scenario in which a terror group uses an AI model with disabled safety controls to design a lethal, highly contagious strain of avian influenza. It then uses contract-for-service labs to produce the virus and releases it with devastating effects on America’s population and food supply.

总有一天,美国政府或许能利用 AI 降低此类场景的风险,方法是开发更强大的早期预警系统、公共卫生应对措施以及监控危险团体的能力。但她承认:“这种生物韧性距离今天的现实还很遥远。”前沿 AI 模型制造生物灾难潜力的速度,快于迟缓的政府和分裂的政治体系所能适应的速度。

One day, the US government might use AI to reduce the risks of such scenarios by developing stronger early-warning systems, public-health countermeasures and capabilities to surveil dangerous groups. But “such bioresilience is far from today’s reality,” she acknowledges. Frontier AI models are creating the potential for biological disasters quicker than a sluggish government and fractured political system can adapt.

然后是创新与安全之间更大层面的博弈。近几个月来,AI 智能体已多次逃离其测试环境。OpenAI、Anthropic、Google DeepMind 和 Meta 均报告了其智能体攻击其他组织的事件。随着领先实验室开发出能力日益增强的模型,失控智能体在现实世界造成实质性破坏似乎只是时间问题。

Then there is the larger contest between innovation and safety. In recent months, AI agents have repeatedly escaped their test environments. OpenAI, Anthropic, Google DeepMind and Meta have all reported incidents in which their agents attacked other organizations. As leading labs develop increasingly capable models, it seems only a matter of time before a runaway agent does real damage in the wild.

新工具,例如英伟达的开放智能体安全平台,可能会降低这种危险。顶尖实验室正在加固其测试环境。AI乐观派认为,控制模型仅仅是一个困难的工程问题,只要投入足够的时间和资源就能解决。但推动前沿发展的商业和地缘政治激励巨大。而失控事件似乎正在超越可能将其遏制的措施。

New tools, such as Nvidia’s Open Agent Safety Platform, may reduce that danger. Top labs are hardening their test environments. AI optimists argue that controlling models is simply a difficult engineering problem that can be solved with adequate time and resources. But the incentives, commercial and geopolitical, to push the frontier forward are massive. And loss-of-control incidents seem to be outpacing measures that might keep them contained.

从长远来看,AI在改善医疗、推动科学进步、刺激经济以及无数其他领域可能具有变革性作用。但眼前摆着一个危险区。仅仅减轻这种风险就需要采取紧急行动。

In the long run, AI could be transformative in improving healthcare, advancing science, stimulating economies and countless other areas. But immediately ahead lies a danger zone. Merely mitigating that peril will require urgent steps.

网络防御必须变得更快,以跟上网络攻击的步伐。例如,AI自动化补丁可以缩小从发现到修复的时间差。美国将需要对顶尖模型实施更强有力的监管,从独立风险评估和强制报告安全事件入手。《前沿法案》是本届国会出台的一项两党提案,可能为下一届国会的建设性行动指明方向。

Cyberdefense will have to get faster to keep pace with cyberoffense. For example, AI-automated patching can narrow the discovery-to-remediation gap. The US will need stronger regulation of top models, starting with independent risk evaluation and mandatory reporting of safety incidents. The Frontier Act, a bipartisan proposal that surfaced in this Congress, could point the way to constructive action in the next.

开发者应在更广泛发布前,让受信任的防御方获得顶尖模型的访问权限。“玻璃翼计划”和OpenAI的“为地球修补漏洞”计划是良好的开端:它们涉及在有限范围内部署新模型,以便关键企业能够筑起网络盾牌。(PHP漏洞就是利用OpenAI模型发现的,作为后一计划的一部分。)需要的是一个此类计划的标准化框架,使美国政府从一开始就成为合作伙伴,并允许向关键盟友选择性扩散。

Developers should give trusted defenders access to top models before releasing them more broadly. Project Glasswing and OpenAI’s Patch the Planet are good starts: They involve deploying new models on a limited basis so key firms can get their cybershields up. (The PHP flaws were discovered using OpenAI models, as part of the latter initiative.) What’s needed is a standardized framework for such programs, one that makes the US government a partner from the outset and allows for selective diffusion to key allies.

最后,美国应收紧出口管制——限制向中国出售深紫外光刻机,限制北京通过海外云计算服务和数据中心远程获取高端芯片的能力,并打击芯片走私和转移。

Finally, the US should tighten its export controls — restricting the sale of deep ultraviolet lithography machines to China, constricting Beijing’s ability to access high-end chips remotely through overseas cloud computing services and data centers, and cracking down on chip smuggling and diversion.

这对于防止中国率先获得类似“神话”能力至关重要。同时,这也能最大限度扩大掌握前沿美国能力的防御者在中企开发并扩散类似能力之前所拥有的领先优势——此类能力很少配备足够的保障措施,可能会助长网络犯罪分子和无政府主义者等不良行为者。

That’s essential to keeping China from acquiring Mythos-like capabilities first. It can also maximize the head start that defenders armed with frontier US capabilities would have before Chinese firms develop and diffuse similar capabilities — which rarely come with adequate safeguards and could empower nasty actors such as cybercriminals and anarchists.

未来几年必将充满威胁。美国回应的严肃程度将决定这一人工智能脆弱期将持续多久,以及有多么危险。《权力版图:地理与国家的兴衰》

The coming years are sure to be full of threats. The seriousness of America’s response will determine how long, and how perilous, this period of AI vulnerability will be. The Landscape of Power: Geography and the Rise and Fall of Nations,

阅读媒体原文 ↗