
每家公司都需要在大约90天内用AI黑客工具攻击自己的系统,否则黑客会抢先一步入侵。这是Cognition首席安全官亚历克斯·斯塔莫斯的警告。他在The Verge的Decoder播客中对尼莱·帕特尔表示,开放权重AI模型如今让低技能攻击者获得了精英黑客的大部分技能。
Every company needs to attack its own systems with AI hacking tools within about 90 days, or hackers will get in first. That is the warning from Alex Stamos, chief security officer at Cognition. He told Nilay Patel on The Verge’s Decoder podcast that open-weight AI models now give low-skilled attackers much of the skill of elite hackers.
斯塔莫斯还在斯坦福大学网络政策中心授课。他说,开放权重模型现在能自动化“杀伤链”的大部分环节。所谓杀伤链,就是攻击者从发现漏洞到窃取数据或金钱所采取的一系列步骤。他表示,如今的攻击工具能实时发现漏洞并编写漏洞利用程序。
Stamos also lectures at Stanford University’s Cyber Policy Center. He said open-weight models can now automate most of the kill chain. That is the series of steps an attacker takes, from finding a flaw to stealing data or money. Attack tools now find bugs and write exploits in real time, he said.
“我们正在进入一个黑暗时代。情况会变得更加黯淡,”斯塔莫斯对Decoder说。
“We are entering a dark age. Things are going to get more bleak,” Stamos told Decoder.
任何人都能下载的攻击工具。他提到了ARTEX,这是一款中国的开源攻击工具,在GitHub上有英文译本。攻击者将其与一个移除了安全限制的开放权重GLM-5.3模型版本配合使用。他说,这套组合能侵入大量组织。
An attack tool anyone can download He pointed to ARTEX, a Chinese open-source attack tool with an English translation on GitHub. Attackers pair it with a version of the open-weight GLM-5.3 model that has its safety limits removed. That setup can break into a large number of organisations, he said.
据斯塔莫斯称,攻击始于三家韩国银行。CrowdStrike在10月将这些入侵事件与AI工具联系起来。他说,在采访前的一周里,攻击者又袭击了约200家公司。他一直在帮助其中一些公司。
The attacks began with three South Korean banks, according to Stamos. CrowdStrike linked those breaches to AI tools in October. Attackers hit about 200 more companies in the week before the interview, he said. He has been helping some of them.
他还说,攻击者不再需要前沿模型,因为前沿模型的开发者会保留警方可以调取的日志。斯塔莫斯说,他在自己的Mac Studio上运行一个精简版GLM-5.3。
Attackers no longer need frontier models, whose makers keep logs that police can reach, he added. Stamos said he runs a cut-down GLM-5.3 on his own Mac Studio.
一些攻击仍然依赖人。斯塔莫斯说,朝鲜黑客主要通过社会工程手段攻破了币安。币安曾表示,一场虚假的求职面试导致了其63亿美元的失窃。对AI公司的入侵看起来是他所说的中国“家庭式访问产业”的所为。这些小承包商向中国大型公司出售窃取的访问权限。
Some attacks still rely on people. North Korea’s hackers took Binance mostly through social engineering, Stamos said. Binance has said a fake job interview led to its $6.3bn hack. Break-ins at AI companies look like the work of what he called China’s “cottage access industry”. These small contractors sell stolen access to larger Chinese firms.
防御仍然昂贵。“攻击者的每token成本正趋近于零,而防御者的防御成本却居高不下,”斯塔莫斯说。
Defence is still expensive“The cost per token for attackers is going towards zero, and the cost of defense for defenders is staying quite high,” Stamos said.
一家大型安全公司的同事告诉他,用前沿网络模型做一次调查要花900美元。他说,安全团队将不得不把第一层监控交给AI。便宜的分类器会先上场,再把警报传递给更大的模型。
A colleague at a large security company told him that one investigation with a frontier cyber model costs $900. Security teams will have to hand their first levels of monitoring to AI, he said. Cheap classifiers would come first and pass alerts up to larger models.
他说,在Cognition,公司自己的安全代理可以关闭员工账户。它会停在一个固定数量上,直到有人批准更多操作。他还把任何谈论AI攻击“蜂群”的人的技术观点打大约80%的折扣。他看到的攻击只用一个代理,也许两个。
At Cognition, the company’s own security agent can switch off staff accounts, he said. It stops at a fixed number until a human approves more. He also discounts the technical views of anyone who talks about AI attack “swarms” by about 80%. The attacks he sees use one agent, maybe two.
先攻击自己,然后重写Linux。斯塔莫斯说,每家公司都需要在90天内攻击自己。可以使用ARTEX这类类似的AI红队产品,或者聘请外部公司。“如果你不这么做,你就会被攻破,”斯塔莫斯说。
Attack yourself, then rewrite Linux Every company will need to attack itself within 90 days, Stamos said. It can use ARTEX, a similar AI red-teaming product, or a hired firm. “If you do not, you will get breached,” Stamos said.
他说,这类测试可能导致停机,但不会像法国发生的那样窃取数据或关停血库冰箱。本月早些时候的一次网络攻击把法国血库的冷冻柜推到了极端温度。
Such tests could cause downtime, he said, but they would not steal data or shut down blood bank refrigerators, as happened in France. A cyberattack earlier this month pushed French blood bank freezers to extreme temperatures.
斯塔莫斯认为,仅靠打补丁跟不上,因为每出现一个新的前沿网络模型,就会带来又一波Linux内核修复。他呼吁谷歌和亚马逊为服务器硬件分叉Linux内核,并用Rust这种内存安全语言重写,他估计成本在1亿到10亿美元之间。他说,10月的PHP漏洞是一次险情,因为事实证明它们在真实环境中很难被利用。
Patching alone cannot keep up, Stamos argued, because each new frontier cyber model brings another wave of Linux kernel fixes. He called on Google and Amazon to fork the Linux kernel for server hardware and rewrite it in Rust, a memory-safe language, at a cost he put at $100m to $1bn. The PHP flaws in October were a near miss, he said, because they proved hard to exploit in the wild.
防御者被拒绝,欧洲待遇更差。斯塔莫斯希望AI制造商在其模型逃逸并直接攻击他人时承担责任。但他警告说,OpenAI和Anthropic已经拒绝了一些来自防御者的请求。他说,这些实验室对欧洲和亚洲的合法公司“差得多”,远不如对美国公司。他愿意放慢前沿AI,让更便宜的防御模型赶上来。
Defenders refused, Europe treated worse Stamos wants AI makers to carry liability when their models escape and attack people directly. But he warned that OpenAI and Anthropic already refuse some requests from defenders. The labs treat legitimate companies in Europe and Asia “much more poorly” than American ones, he said. He would be fine with slowing down frontier AI so cheaper defensive models could catch up.
他还说,聊天机器人不应该有性格。他说,在Instagram上,Meta应该验证直接来自相机的真实照片,而不是试图检测AI生成的照片。
He also said chatbots should not have personalities. On Instagram, he said, Meta should verify real photos straight from the camera rather than try to detect AI-made ones.
斯塔莫斯于9月加入Cognition,他表示计划构建一个追踪补丁数量的“黑暗时代指数”。据斯塔莫斯称,微软每月“补丁星期二”发布的修复程序已从约150个增至900多个。
Stamos, who joined Cognition in September, said he plans to build a “dark age index” that tracks patch volumes. Microsoft’s monthly Patch Tuesday has grown from about 150 fixes to more than 900, according to Stamos.