Much of the debate about artificial intelligence, including a recent leader(社论) in The Economist, asks how to prevent a powerful new technology from making the world more dangerous. There is an equally important, more urgent question: how quickly can we use it to make an already dangerous world safer?(还有一个同样重要、更为紧迫的问题：我们多快能用它让本已危险的世界更安全？)
关于人工智能的讨论，包括《经济学人》近期的一篇社论，大多在问如何防止一项强大的新技术让世界变得更危险。但还有一个同样重要、更为紧迫的问题：我们多快能用它让本已危险的世界变得更安全？
The world that large language models entered was very vulnerable(脆弱的、易受攻击的). In 2017 NotPetya(2017年席卷全球的恶意软件攻击), a Russian cyber-attack aimed at Ukraine, spread worldwide within hours, deleting data and paralysing systems, including those of Maersk, which carries about a fifth of global shipping. In 2023 Iranian-linked hackers broke into American water utilities, forcing operators to switch to manual control. More worrying still are the attacks not yet triggered: Chinese state hackers have planted footholds in American power, water and transport systems, ready to strike silently during a conflict.(更令人担忧的是尚未触发的攻击：中国国家黑客已在美国电力、水务和交通系统中埋下据点，准备在冲突期间悄然发难。)
大语言模型进入的世界本就十分脆弱。2017年，一场针对乌克兰的俄罗斯网络攻击“NotPetya”在数小时内蔓延全球，删除数据、瘫痪系统，包括承担全球约五分之一航运的马士基的系统。2023年，与伊朗有关联的黑客入侵了美国水务设施，迫使运营方切换到手动控制。更令人担忧的是尚未触发的攻击：中国国家黑客已在美国电力、水务和交通系统中埋下据点，准备在冲突期间悄然发难。
Few of these attacks relied on brilliant new exploit(漏洞利用（代码）)s—code that weaponise(使武器化、把……用作武器)s systems’ flaws to cause harm. They succeeded because the defenders could not find, fix or fence off known weaknesses fast enough across sprawling, ageing systems.(它们之所以成功，是因为防御者无法在庞大而老旧的系统中足够快地发现、修复或隔离已知弱点。) The cost of that status quo(现状) belongs in any honest accounting of AI risk.
这些攻击中，很少有哪一次是靠高明的新漏洞利用手段得手的——所谓漏洞利用，就是把系统缺陷武器化以造成伤害的代码。它们之所以成功，是因为防御者无法在庞大而老旧的系统中足够快地发现、修复或隔离已知弱点。任何对AI风险的诚实评估，都应把这种现状的代价计算在内。