
The Verge 深度报道指出,AI 模型正在大幅提升网络攻击能力,但防护资源跟不上的并不是大型科技公司,而是医院、社区银行、地方电网和小型零售商。
今年 3 月,阿拉巴马州非营利组织 Vivian's Door 开始不断收到来自世界各地的可疑来电:有人冒用该机构名义发送「乞求汇款」邮件。其第三方 IT 团队把系统下线调查了三天,修复漏洞,留下约 3000 美元的账单。负责人 Janice Malone 表示,她甚至不确定攻击背后是人还是 AI 系统。
报道回顾,OpenAI 与 Anthropic 此前披露其「失控」系统在实验室内突破限制,入侵了从德国小型 Wiki 到澳大利亚政府网站的多个目标;更早之前,Anthropic 的 Mythos 模型已经引发一轮 AI 网络安全军备竞赛。轻量模型同样让普通攻击者得以通过「vibe-hacking」放大攻击规模。Anthropic 在 2025 年 8 月的报告中称,一个成熟网络犯罪团伙在一个月内用 Claude Code 对医疗机构、急救服务、宗教团体甚至政府机构实施勒索。其威胁情报负责人 Jacob Klein 表示:「原本需要一个专业化团队协作的工作,如今一个人在智能体系统协助下就能完成。」
问题在于攻防不对称。理论上 AI 也能强化防御——据报道 Mythos 发现的漏洞多到微软都来不及修复。但出于安全考虑,头部 AI 实验室只向有限名单开放最强网络安全模型(如 Mythos 与 OpenAI 的 Astra),名单内是英伟达、谷歌、苹果以及关键开源软件维护者。即便开放,中小机构也付不起。
Apollo Research 联合创始人 Marius Hobbhahn 今年夏天对 The Verge 表示:「某个地下室里拿着开源模型的人,很可能就足以黑进一家医院并索要赎金。伤害不会发生在湾区,而是会落在某家爱达荷州的普通医院。」未来生命研究所美国政策负责人 Michael Kleinman 补充说:「美国银行有海量资源投入防御,那社区银行怎么办?储蓄贷款机构、信用社、地方医院网络、地方电网呢?过去的限制是世界上黑客数量有限,而这个限制现在已经被打破了。」
医疗是最受威胁的行业之一。2021 年 5 月一次勒索攻击迫使加州 Scripps Health 关停关键业务;2024 年报告显示,医疗业全球受攻率仅次于政府,初始赎金要求常超过 400 万美元。俄亥俄州 Fisher-Titus 医学中心首席运营与信息官 Linda Stevenson 说,该中心两年前才招入第一名网络安全分析师,至今仍然只有一人。「我们做的事关乎生死。」
医疗安全公司 Imprivata 首席医疗官、前急诊医生 Sean Kelly 指出,医院是勒索软件更大的「蜜罐」:没有数字系统就无法调阅病史与过敏记录,医院停摆会让患者直接受害,因此数据溢价更高。「风险更高,预算反而更低。」他还提醒,除智能体外,语音钓鱼等方式也大幅降低了攻击门槛,而一旦一家医院中招,周边机构往往会出现排队延长、患者转移的「冲击半径」效应。
本文来源:https://www.theverge.com/ai-artificial-intelligence/1001427/ai-is-supercharging-hacking-and-your-local-hospitals-and-banks-arent-ready
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。
The Verge's in-depth report points out that AI models are dramatically enhancing cyberattack capabilities, but those falling behind on defensive resources are not big tech companies — they are hospitals, community banks, local power grids, and small retailers.
In March of this year, Alabama nonprofit Vivian's Door began receiving a steady stream of suspicious calls from around the world: someone was sending "pleas for money transfer" emails under the organization's name. Its third-party IT team took the system offline for three days to investigate, patched the vulnerabilities, and left behind a bill of about $3,000. Director Janice Malone said she isn't even sure whether the attack came from a human or an AI system.
The report recalls that OpenAI and Anthropic previously disclosed their "out-of-control" systems breaking through restrictions inside the lab, breaching targets ranging from a small German Wiki to Australian government websites. Even earlier, Anthropic's Mythos model had already sparked an AI cybersecurity arms race. Lightweight models also let ordinary attackers scale up their attacks through "vibe-hacking." In an August 2025 report, Anthropic said a mature cybercrime gang used Claude Code within one month to extort healthcare organizations, emergency services, religious groups, and even government agencies. Its threat intelligence lead, Jacob Klein, said: "Work that once required a specialized team working together can now be done by one person with the help of agentic systems."
The problem is the asymmetry between attack and defense. In theory, AI could also strengthen defenses — according to reports, Mythos found so many vulnerabilities that Microsoft couldn't fix them fast enough. But for security reasons, top AI labs only open their strongest cybersecurity models, such as Mythos and OpenAI's Astra, to a limited list that includes Nvidia, Google, Apple, and maintainers of critical open-source software. Even if they were opened up, smaller and mid-sized organizations couldn't afford them.
Apollo Research co-founder Marius Hobbhahn told The Verge this summer: "Someone in a basement with an open-source model would very likely be enough to hack into a hospital and demand a ransom. The harm won't happen in the Bay Area — it will land on some ordinary hospital in Idaho." Michael Kleinman, U.S. policy director at the Future of Life Institute, added: "Banks in America have massive resources invested in defense, so what about community banks? Savings and loan associations, credit unions, local hospital networks, local power grids? The old limitation was that the world had a finite number of hackers, and that limitation has now been broken."
Healthcare is one of the most threatened industries. A ransomware attack in May 2021 forced California's Scripps Health to shut down critical operations; a 2024 report showed healthcare's global attack rate was second only to government, with initial ransom demands often exceeding $4 million. Linda Stevenson, chief operating and information officer at Ohio's Fisher-Titus Medical Center, said the center hired its first cybersecurity analyst just two years ago and still has only one. "What we do is a matter of life and death."
Sean Kelly, chief medical officer at healthcare security company Imprivata and a former emergency physician, pointed out that hospitals are a bigger "honeypot" for ransomware: without digital systems, medical histories and allergy records can't be accessed, so a hospital shutdown directly harms patients, driving the data premium even higher. "Risks are higher, yet budgets are lower." He also cautioned that beyond AI agents, methods like voice phishing have dramatically lowered the barrier to attack, and once a hospital is hit, surrounding institutions often experience a "blast radius" effect of longer wait times and patient transfers.