这本该是“网络安全意识月”。专家现在称之为“黑色十月”。近几周来，工程师们必须不断遏制的计算机黑客攻击和漏洞的正常流量已变成洪流。原因是人工智能，它给网络犯罪加了速。开源AI模型使攻击者能够以非人的效率、能力和持久性开展工作。结果，多年积累的关键漏洞在几天内就被暴露出来。
It was meant to be “cyber-security awareness month”. Experts now call it “Black October”. In recent weeks the normal flow of computer hacks and bugs that engineers must constantly stem has become a torrent. The cause is artificial intelligence, which has turbocharged cyber malfeasance. Open-source AI models have empowered attackers to work with inhuman efficiency, competence and persistence. As a result, years’ worth of critical flaws have been exposed in days.
这些漏洞从令人震惊到荒谬可笑不等。工程师们匆忙修复了广泛使用的编程语言PHP中的一个漏洞，该漏洞允许攻击者接管任何使用它的计算机。欧洲各地超市的冰柜按照网络犯罪分子的指令断电。而一个关键的互联网安全证书被发现其私钥为“4”，而它本应有上百位数字长。没有人知道在私下里可能已发现并利用了哪些漏洞。
Those flaws range from shocking to ridiculous. Engineers have rushed to fix a flaw in PHP, a widely used programming language, that let attackers take over any computer using it. Freezers in supermarkets across Europe powered down on the instructions of cyber-criminals. And a crucial internet security certificate was revealed to have a private key of “4”, when it should have been a hundred digits long. Nobody knows what bugs might have been found and exploited in private.
这些事件让许多评论人士（包括本刊）感到警醒，他们曾希望AI对网络安全的冲击是短暂的，并且这项技术本身对数字系统防御者的帮助会大于对犯罪分子的助长。这些希望正在破灭。相反，现代生活所依赖的数字基础设施在AI面前显得越来越脆弱。需要采取紧急行动来保卫它。
These events have been sobering for the many commentators, including this newspaper, who hoped that the disruption to cyber-security from ai would be short-lived, and that the technology itself would help those who defend digital systems more than it would empower crooks. Those hopes are being dashed. Instead, the digital infrastructure on which modern life depends looks increasingly vulnerable to ai. Urgent action is needed to defend it.
为什么防御者陷入困境？第一个原因是他们中的许多人没有最先进的工具。今年4月，领先的模型开发商Anthropic推出了“玻璃翼计划”（Project Glasswing），这是一个仅限受邀者加入的俱乐部，可以使用其前沿模型的全部网络能力。它原本打算广泛扩大成员范围。OpenAI的一个类似项目在启动时承诺使其工具的使用民主化。
Why are defenders struggling? The first reason is that many of them do not have the most advanced tools. In April Anthropic, a leading model-maker, launched “Project Glasswing”, an invitation-only club with access to the full cyber capabilities of its frontier models. It originally intended to broaden the membership widely. A similar programme from OpenAI was launched with a promise to democratise access to its tools.
然而，获取强大AI已成为一个地缘政治问题。这两家模型制造商现在都需要获得美国联邦政府的批准，才能接纳新成员加入它们的俱乐部，因此扩大成员规模已成为一个痛苦的政治过程。与此同时，门外的人必须用与驱动攻击相同的开源工具来保护自己。有些人陷入了无人区：无法获得美国技术，但又害怕使用中国替代品。
However, access to powerful AI has become a geopolitical issue. Both of the model-makers now need sign-off from America’s federal government before admitting new members to their clubs, so expanding them has become a painful political process. In the meantime those outside the gates must defend themselves with the same open-source tools that are powering the attacks. Some are stuck in a no-man’s land: unable to access American tech, but also scared of using Chinese alternatives.
第二个问题是，当AI被善意使用时，它会造成友军误伤问题。针对新发现的漏洞的“补丁”一经发布，黑客的AI模型就可以检查它。这可能揭示出弱点究竟是什么，以及如何在尚未修复的系统中利用它。如果防御者真的不走运——就像他们在php问题上那样——补丁可能会在发布给防御者用于修复代码之前就泄露。攻击者随后就可以对仍然毫无防御能力的系统发起攻击。
The second problem is that when AI is used benignly it creates a friendly-fire problem. The moment that a “patch” for a newly identified vulnerability is released, hackers’ AI models can examine it. This may reveal exactly what the weakness was and how to exploit it in systems that have not yet been fixed. If defenders are really unlucky—as they were with php—the patch may leak before it is released for defenders to use to fix their code. Attackers can then operate against still-defenceless systems.
最后一个问题是，攻击者可以更快、更大规模地利用平淡无奇的漏洞（见商业版块）。网络安全不仅仅在于理论上拥有强大的系统，还在于避免实践中的松懈。不幸的是，AI让人类的疏忽更容易被利用，却没有让计算机用户变得更明智。一台联网的冰箱运行过时软件或使用默认密码，几十年来一直是安全隐患。然而，直到现在，攻击者才发现如此容易地自动化搜索和利用弱点，并行运行数千次攻击。
A final issue is that attackers can exploit prosaic vulnerabilities faster and on a greater scale (see Business section). Cyber-security is not just about having strong systems available in theory, but avoiding laxity in practice. Unfortunately, AI has made human carelessness easier to exploit without making computer users any wiser. An internet-connected fridge running outdated software or using a default password has been a security hazard for decades. Only now, however, have attackers found it so easy to automate the search for and exploitation of weak points, running thousands of attacks in parallel.
企业和政府必须果断采取行动应对这些威胁。制造网络武器的不平衡至关重要。在火力对等的情况下，防御者依赖一项狭窄的天然优势：他们只需找到并修补单个漏洞——这些孤立的问题理论上可能有助于形成突破口。攻击者通常需要掌握一整条漏洞链才能侵入系统。
Companies and governments must act decisively against these threats. Generating an imbalance in cyber-weaponry is crucial. With equal firepower, defenders rely on a narrow natural advantage: they must find and patch only single vulnerabilities—isolated problems which might in theory help create a breach. Attackers typically need to know of a whole chain of those vulnerabilities in order to breach a system.
如今，AI驱动的攻击正在激增，这一优势过于狭窄。防御者需要更强大的武器库。美国政府应对网络防御计划的成员资格采取更宽松的态度，或者至少加快批准新加入者。如果攻击者的能力被进一步压制在前沿之后，情况也会更好。中国政府已经认识到向所有人分发强大网络武器的愚蠢之处，并正在要求实验室限制新发布系统的破坏潜力。
Now that AI-powered attacks are proliferating, this advantage is too narrow. Defenders need a bigger armoury. America’s government should take a more permissive attitude to membership of cyber-defence programmes, or at least approve new joiners faster. It would also be better if attackers’ capabilities were pushed further behind the frontier. China’s government has recognised the folly of distributing potent cyber-weapons to all, and is requiring labs to limit the destructive potential of newly released systems.
尚未持续使用AI来发现和封堵安全漏洞的组织已经落后，必须迎头赶上。一些实体需要更多帮助。在美国，得克萨斯州“流域250项目”等举措正在向基础设施提供商输送资源，以帮助解决基础问题。这些努力应当扩大。假以时日，软件行业或许能够对某些代码进行形式化验证，从而证明其无法被黑客攻破，就像数学家证明定理一样。
Organisations that are not yet using AI constantly to find and close security holes have fallen behind the curve and must catch up. Some entities need more help. In America efforts like Texas’s “Project Watershed 250” are funnelling resources to providers of infrastructure to help with the basics. Those efforts should be expanded. In time, the software industry may be able to verify some code formally, thereby proving that it is unhackable, much as a mathematician proves a theorem.
商业、基础设施和人际关系已经依赖于互联网是安全的这一假设。因此，网络安全的突然丧失将对社会构成严重威胁。过去一个月里，网络安全意识有所提高。行动必须跟上。■
Commerce, infrastructure and human relationships have come to rely on the assumption that the internet is secure. A sudden loss of cyber-security would therefore pose a grave threat to society. Cyber-security awareness has risen over the past month. Action must follow. ■