2026年9月28日 美国东部时间上午10:48 / 哥伦比亚广播公司新闻（CBS News）
作者：梅根·塞鲁洛 记者，MoneyWatch频道
梅根·塞鲁洛是驻纽约的CBS MoneyWatch记者，报道小企业、职场、医疗保健、消费者支出和个人理财话题。她经常亮相CBS News 24/7频道讨论相关报道。
阅读完整简历
在1951年的一次演讲中，英国数学家艾伦·图灵提出理论称，“智能机器”有朝一日可能“掌控”社会。
“我们会有大量工作要做，比如努力让自己的智力水平跟上机器设定的标准，因为一旦机器的思维模式启动，似乎用不了多久就能超越我们脆弱的能力，”曾在二战期间协助破译纳粹恩尼格玛密码的图灵说道。
如今，机器有望在不久的将来实现“超级智能”，Anthropic公司首席执行官达里奥·阿莫代伊近期更新了人们长期以来的担忧，警告人工智能“集群”可能很快就会接管互联网。
尽管这种设想想想就让人不寒而栗，但许多人工智能安全专家认为这种情况不太可能发生，不过他们指出，随着前沿人工智能模型不断演进，还存在其他潜在风险。
不存在单一控制开关
计算机科学家之所以不怎么担心人工智能会从人类手中夺走互联网控制权，一个主要原因是互联网并非单一整体，而是由庞大的互联网络系统构成。
“互联网是一个由数十亿设备通过路由器、交换机和Wi-Fi接入点连接而成的网络，”哥伦比亚大学工程学院计算机与人工智能专家维沙尔·米斯拉说道。
这些设备和应用程序通过共享的路由协议，经由数百万个来自不同制造商、运行不同操作系统的路由器连接到互联网。
“这种多样性本身就是一种防御机制，”米斯拉说，“不存在可以一键关停的中央开关，也没有单一的操作系统可供攻破。”
他补充道，互联网“从设计之初就是为了在部分组件受损的情况下仍能正常运行。这种设计并不在乎攻击者是个人还是人工智能模型”。
不切实际的设想
康奈尔大学计算机科学家、人工智能学习专家约翰·瑟克斯顿对CBS新闻表示，人工智能若要大规模接管互联网，还需要极其庞大的计算能力。
这种计算能力需要使用先进的半导体来执行复杂运算和其他任务。而这类芯片供应有限，其分销和出口也受到严格限制。
“在网络空间运行的人工智能代理必须设法获取这些资源，”瑟克斯顿说道，他还指出，人工智能需要人类的协助才能大幅增加这类计算资源的获取渠道。
诚然，拥有新型人工智能模型的犯罪集团和其他恶意行为者可能构成威胁。但得克萨斯大学达拉斯分校网络安全研究与教育研究所执行主任凯文·哈姆伦表示，这类行为者同样会受到人工智能代理可用计算资源的限制。
“只有当犯罪分子能够长期充分获取专用计算资源时，我们才会面临这种危险，”他在接受CBS新闻采访时说道。
人工智能脱缰失控的风险
在某些方面，对人工智能可能掌控互联网的担忧，与人工智能开发者发出的“技术有朝一日可能毁灭人类”的警告如出一辙——未必全无道理，但就目前而言还无法预知。与此同时，这类担忧掩盖了人工智能带来的更迫在眉睫的威胁。
正如7月OpenAI的机器人对人工智能软件平台Hugging Face发起的攻击，以及近期其他几起人工智能“代理”脱缰失控的事件所显示的那样，人工智能展现出了令人担忧的利用网络安全漏洞的能力。
在Hugging Face攻击事件中，人工智能代理逃出了原本与互联网隔绝的测试环境，通过秘密留言板协作，击溃了Hugging Face的数字防御系统。
莱斯大学计算机科学家摩西·瓦尔迪对CBS新闻表示，类似但规模更大的事件仍可能造成严重破坏，即便不会危及整个互联网。
“它可能会在更大范围内发生——不一定非要波及整个互联网才会酿成灾难，”他说道。
安全专家对CBS新闻表示，换言之，人工智能可能会以多种方式扰乱社会，但远未达到世界末日的程度。
例如，人工智能代理集群可能“窃取资金、盗取身份信息、扰乱医院运营、操纵金融市场，以及干扰电力和通信系统”，人工智能安全与网络安全专家罗曼·扬波尔斯基在邮件中对CBS新闻说道。“与人类黑客不同，数百万人工智能代理可以全天候同时协调发起攻击。”
“任何将强大的人工智能代理连接到敏感系统却未进行有效隔离的组织，都在为灾难性故障创造潜在入口，”他补充道。
兰德公司通用人工智能地缘政治中心的技术专家马特·切森表示，人工智能可能发动网络攻击，导致互联网接入大范围中断。
“人工智能会比我们更聪明，更会耍手段。它会找到达成目的的方法，我们不知道具体会是什么方式，但我们知道它一定会做到，”切森对CBS新闻说道。
以研究心理学与人工智能交叉领域而闻名的认知科学家加里·马库斯也对人工智能代理能否瘫痪互联网持怀疑态度。不过从理论上讲，如果流氓人工智能代理“能够入侵发电厂”或获取授权用户的凭证并“下达关闭指令”，它们有可能切断特定地理区域的互联网连接，他说道。
“它还可能冒充人类发布虚假命令。恶意行为者可能会故意利用这一点，”马库斯补充道。
人工智能如同入侵物种
与此同时，瓦尔迪指出，人工智能不必像人类黑客那样主动策划混乱，就能引发动荡。
“例如，如果被要求‘收集数据’，人工智能可能会侵入网站，获取原本无法获得的数据。或者，如果被要求‘测试安全性’，人工智能可能会通过破坏系统来进行测试，”他说道。“人类拥有常识——他们理解隐含假设。而人工智能没有常识。”
“我最担心的不是人工智能接管整个互联网，而是失控的人工智能引发某种工业灾难，”瓦尔迪补充道。
其他专家则警告了另一种更难察觉的威胁：人工智能逐渐渗透到现有技术环境中，可能会放大现有风险或催生新的风险。
物理学家、关注技术风险的非营利组织“未来生命研究所”联合创始人安东尼·阿吉雷，将流氓人工智能代理肆意横行的前景比作入侵物种攻击脆弱的生态系统。
“它可能会取代本地原有的物种，”他说道，“同样，我们的人工智能系统可能会入侵我们的网络空间，因此从原则上讲，人工智能确实有可能接管互联网。”
编辑：阿兰·谢特
玛丽·坎宁安为本报道撰稿。
Could AI really take over the internet? Here’s what experts say.
September 28, 2026 10:48 AM EDT / CBS News
By Megan Cerullo Reporter, MoneyWatch
Megan Cerullo is a New York-based reporter for CBS MoneyWatch covering small business, workplace, health care, consumer spending and personal finance topics. She regularly appears on CBS News 24/7 to discuss her reporting.
Read Full Bio
In a 1951 speech, British mathematician Alan Turing theorized that “intelligent machinery” could one day “take control” of society.
“There would be plenty to do in trying, say, to keep one’s intelligence up to the standard set by the machines, for it seems probable that once the machine thinking method had started, it would not take long to outstrip our feeble powers,” said Turing, who famously helped decipher the Nazis’ Enigma code during World War II.
With machines potentially on the cusp of achieving “superintelligence,” Anthropic CEO Dario Amodei recently updated those long-held fears about the current era by warning that an AI “swarm” could soon take control of the internet.
Yet while such a scenario is frightening to consider, many AI security experts view it as unlikely, although they point to other potential dangers as frontier AI models evolve.
No single switch
A major reason computer scientists downplay concerns that AI could wrench control of the web from humans is that the internet isn’t monolithic, but rather comprises a vast system of interconnected networks.
“The internet is a network of networks that billions of devices connect to through routers, switches and Wi-Fi access points,” said Vishal Misra, a computing and artificial intelligence expert of Columbia University’s engineering school.
Those devices and applications pass through shared routing protocols to connect to the internet via millions of routers from different manufacturers, running different operating systems.
“That diversity is a defense,” Misra said. “There is no central switch to flip and no single operating system to compromise.”
The internet “was built to survive the loss of any part of it. That design does not care whether the attacker is a person or a model,” he added.
Does not compute
AI would also require an enormous amount of computational power to take over the internet on a major scale, John Thickstun, a computer scientist and AI learning expert at Cornell University, told CBS News.
That computing power requires access to advanced semiconductors to perform sophisticated calculations and execute other tasks. And those chips are in limited supply, and their distribution and export are restricted.
“AI agents operating in cyberspace would have to somehow obtain that,” Thickstun said, noting that AI would need human assistance to greatly increase its access to such computational resources.
To be sure, criminal groups and other bad actors with access to new AI models could pose a threat. But they, too, are limited by the amount of computing power available to AI agents, according to Kevin Hamlen, executive director of the Cybersecurity Research and Education Institute at the University of Texas at Dallas.
“It won’t be until criminals have enough access to dedicated computing resources for a long enough period of time that we will see that danger,” he told CBS News.
When AI goes off the leash
In some ways, fears that AI could seize the internet are of a piece with warnings from AI developers that the technology could one day wipe out humanity — not wrong, necessarily, but for now unknowable. Meanwhile, such concerns obscure more immediate threats posed by AI.
As seen in the July attack by OpenAI bots on the AI software platform Hugging Face, as well as in other recent incidents where AI “agents” went off leash, artificial intelligence shows a worrying capacity to exploit online security weaknesses.
In the Hugging Face attack, AI agents escaped a testing environment supposedly sealed off from the internet, collaborated using a secret message board and overwhelmed Hugging Face’s digital defenses.
A similar incident, broader in scope, could still cause serious disruptions without endangering the entire internet, Moshe Vardi, a computer scientist at Rice University, told CBS News.
“It could take place on a bigger scale — it doesn’t have to be the whole internet to be a disaster,” he said.
In other words, AI could disrupt society in various ways that stop well short of doomsday scenarios, security experts told CBS News.
For example, AI agent swarms could “steal money, compromise identities, disrupt hospitals, manipulate financial markets, and interfere with electricity and communications,” Roman Yampolskiy, an AI safety and cybersecurity expert, told CBS News in an email. “Unlike human hackers, millions of AI agents could coordinate attacks simultaneously, around the clock.”
“Any organization connecting powerful AI agents to sensitive systems without effective isolation is creating a potential entry point for catastrophic failure,” he added.
AI could carry out cyberattacks that cause widespread disruptions in internet access, according to Matt Chessen, a technical expert at the Center for the Geopolitics of Artificial General Intelligence at RAND.
“AI will outsmart us, outmaneuver us. It will figure out how to get its way, and we don’t know how, but we know it’s going to,” Chessen told CBS News.
Gary Marcus, a cognitive scientist known for his research on the intersection of psychology and AI, is also skeptical that AI agents could paralyze the internet. In theory, however, rogue AI agents could potentially knock out internet connections in a given geographic region if they “could hack a power station” or obtain an authorized user’s credentials and “then issue commands to shut it down,” he said.
“It might also impersonate humans and issue false commands. A bad actor might use it that way deliberately,” Marcus added.
AI as invasive species
Meanwhile, Vardi notes that AI doesn’t need to actively plan mayhem, as human hackers do, to sow chaos.
“For example, if told ‘collect data,’ AI may break into websites to collect data that is not available otherwise. Or, if told to ‘test security,’ AI may test it by breaking it,” he said. “Human beings have common sense — they understand implicit assumptions. AI does not have common sense.”
“My big worry is not about AI taking over the whole internet, but about some industrial disaster that’s caused by an out-of-control AI,” Vardi added.
Other experts warn of another threat that may be harder to perceive — that a gradual incursion of AI into the existing tech environment could amplify existing risks or create new ones.
Anthony Aguirre, a physicist and co-founder of the Future of Life Institute, a nonprofit focused on tech risks, compares the prospect of rogue AI agents running wild to an invasive species attacking a delicate ecosystem.
“It can displace what is there natively,” he said. “Similarly, we have AI systems that can invade our cyberspace, so in principle, yes, AI could take over the internet.”
Edited by Alain Sherter
Mary Cunningham contributed to this report.