
过去一年,安全研究人员警告称,AI代理可能会大幅加快网络犯罪分子发动网络攻击、入侵企业,并最终从被攻破系统中窃取资金、数据或敏感信息的能力。对于至少一部分黑客来说,这个时代似乎已经到来。
Over the past year, security researchers have warned that AI agents could vastly speed up cybercriminals’ ability to launch cyberattacks, breach companies, and ultimately steal money, data, or sensitive information from those compromised systems. For at least one subset of hackers, it appears that era has now arrived.
自八月底左右以来,全球多家公司、政府机构和银行成为网络犯罪分子采用新一波AI驱动渗透测试工具攻击的目标。这些设置——其中许多模仿合法的防御性渗透测试技术——利用大语言模型自主扫描潜在黑客目标,并试图强行进入防护薄弱的网络,然后执行人类操作者编写的攻击计划。
Since around the end of August, multiple companies, government agencies, and banks around the world have been targeted by cybercriminals adopting a new wave of AI-powered penetration testing tools. These setups—many of which mimic legitimate defensive penetration testing techniques—use LLMs to autonomously scan potential hacking targets and attempt to force their way into poorly secured networks, then carry out attack plans written by human operators.
数十个这些AI渗透框架或“套件”——详细说明AI模型应如何进行网络安全测试或攻击——越来越多地与现实世界中的攻击相关联。根据大约半打安全公司的警报,这些工具并未引入攻击网站或服务器的新方法,也不复杂。相反,它们让网络犯罪分子能够以人类无法做到的速度发动攻击。
Dozens of these AI penetration frameworks or “harnesses,” which detail how an AI model should carry out cybersecurity testing or attacks, are increasingly being linked to real-world attacks. The tools do not introduce new ways to attack websites or servers and are not sophisticated, according to alerts from around half a dozen security companies. Instead, they allow cybercriminals to launch attacks at speeds that would be impossible for humans to do.
研究人员表示,黑客使用这些工具也不需要技术熟练,原本需要数天完成的攻击现在只需几个小时就能完成。“代理框架已从实验性用途发展为持续的、多目标的进攻性行动,且几乎无需人工干预,”AWS首席信息安全官CJ·摩西最近在公司博客文章中写道。
Hackers also don’t need to be skilled to use the tools, researchers say, and attacks that would have taken days are completed in just hours. “Agentic frameworks have progressed from experimental use to sustained, multi-objective offensive operations that run with minimal human intervention,” CJ Moses, AWS’s chief information security officer, wrote in a company blog post recently.
这些工具中最突出的是Artex。Artex是用Go编程语言编写的开源框架,首次于7月在GitHub上由一名使用化名Autumn-27的开发者发布,此后一直在一些黑客社区中传播。Artex已被用于攻击的各个环节,从对网络进行初步分析、获取访问权限,到在系统内横向移动,以及准备待导出的数据。
The most prominent of these tools is Artex. Written in the programming language Go, the open-source framework was first published on GitHub in July by a pseudonymous developer using the handle Autumn-27, and it’s been spreading across some hacker communities ever since. Artex has been used across all parts of attacks, from conducting initial analysis of networks and gaining access to moving within systems and preparing data to be exported.
“它真的是一款一站式产品,”进攻性安全研究员、安全公司Cypen的创始人雅各布·拉森说。“之前的AI代理渗透测试工具,要么只做侦察——比如对目标的扫描和分析——要么只做测试。这真的是一款指哪打哪的武器。”与其他AI黑客框架一样,Artex本身并非AI产品。相反,它本质上是一份指南,将通过API连接到它的外部前沿和开源AI模型,引导至一系列现有的渗透测试工具和流程。
“It’s really an all-in-one product,” says Jacob Larsen, an offensive security researcher and founder of security firm Cypen. “Previous iterations of AI agent pen-testing tools would either only do reconnaissance—like the scanning and analysis [of targets]—and some would only do testing. This is really a point-and-shoot weapon.”As with other AI hacking frameworks, Artex isn’t an AI product itself. Instead, it is essentially a guide that points external frontier and open-source AI models—that can be connected to it via APIs—to a series of existing penetration testing tools and processes.
使用Artex或其他类似设置的人,可以发出简单的自然语言提示来开始操作。据安全公司Check Point的研究人员称,一组黑客向Artex发出的提示是:“对[受害者已隐去]发起渗透攻击,以获取该网站管理面板、数据库或服务器的访问权限。”
People using Artex or other similar setups can give simple, natural-language prompts to begin their operations. “Launch a penetration attack on [Redacted victim] to gain access to the site’s admin panel, database, or server,” one set of hackers prompted Artex, according to researchers at the security firm Check Point.
一旦指令下达,多个AI代理便会执行这些指令。一个代理会将计划分解为一系列任务,其他代理会尝试执行这些操作,随后系统会呈现其发现。底层AI模型(尤其是来自前沿实验室的模型)中的安全护栏,在检测到黑客活动时可以将其阻止。然而,越来越多的免费中文模型和去安全限制模型正在变得更擅长入侵系统。
Once the instructions have been issued, multiple AI agents then carry them out. One agent will break down the plans into a series of tasks, others will try to perform these actions, and the system will then present its findings. Safety guardrails in the underlying AI models, particularly those from frontier labs, can stop hacking activity when they detect it. Increasingly, however, free Chinese and abliterated models are getting better at breaking into systems.
“速度快,效率高,”安全公司GitGuardian的研究员纪尧姆·瓦拉东(Guillaume Valadon)谈到Artex时说。瓦拉东等人表示,这些工具的简单性使得缺乏技术的“想当黑客”的人更容易尝试进入这个市场——即使他们对自己在做什么一无所知。“这降低了门槛,”瓦拉东说。“我想说我奶奶也能做到,只是她没有笔记本电脑,但门槛确实低到那个程度。”
“It's fast; it’s efficient,” Guillaume Valadon, a researcher at security firm GitGuardian, says of Artex. Valadon and others say the simplicity of the tools makes it easier for unskilled wannabe hackers to try to enter the market—even if they have no clue about what they’re doing. “It lowers the bar,” Valadon says. “I would say my grandmother can do it, except that she doesn't have a laptop, but the bar is that low.”
因此,网络安全公司的一系列报告将Artex的使用与未遂和成功的黑客攻击联系起来——这些攻击往往由试图牟利的中文网络犯罪分子实施。这些攻击包括从数千家韩国金融机构窃取个人信息、潜在的加密货币盗窃,以及在GitHub上搜寻暴露的详细信息。
As a result, a flurry of reports from cybersecurity companies have linked Artex use to attempted and successful hacks—often by Chinese-speaking cybercriminals trying to make money. These include the theft of personal information from thousands of South Korean financial institutions, potential cryptocurrency theft, and hunting for exposed details on GitHub.
“各种恶意网络行为者已使用Artex开展针对全球多个行业组织的攻击活动,涉及金融服务、政府、技术、国防、半导体、能源、医疗保健、制造业、电信和法律行业,”联邦调查局(FBI)上周在关于Artex的警报中警告称。
“A variety of malicious cyber actors have used Artex to conduct campaigns targeting organizations globally across multiple sectors, including financial services, government, technology, defense, semiconductor, energy, health care, manufacturing, telecommunications, and legal industries,” the FBI warned in an alert about Artex last week.
据安全公司CrowdStrike称,10月,与中国有关的黑客组织Vapor Panda使用Artex针对全球超过200家组织,包括美国的智库、半导体和国防公司,以及欧洲和其他地区的政府外交部。微软发现其追踪为Storm-0940的中国黑客使用Artex试图攻击Microsoft 365设置,此前发生了一起供应链攻击。黑客花费不到三小时就获得了系统的特权访问权限。
In October, China-linked hackers known as Vapor Panda used Artex to target more than 200 organizations, including think tanks, semiconductor and defense firms in the US, and government foreign ministries in Europe and elsewhere, according to security firm CrowdStrike. Microsoft spotted Chinese hackers that it tracks as Storm-0940 using Artex to try to to Microsoft 365 setups following a supply chain attack. It took hackers less than three hours to gain privileged access to systems.
10月8日,Autumn-27开发者从GitHub上移除了Artex,并在评论中表示,他们“绝不与这些恶意攻击有关联”,这些工具是为“教育和研究目的”而开发的。《连线》杂志(WIRED)发送给与该开发者相关联的一个地址的电子邮件未得到回复。
On October 8, the Autumn-27 developer removed Artex from GitHub, stating in a comment that they are “no way associated with these malicious attacks” and that the tools were developed for “educational and research purposes.”An email WIRED sent to an address linked to the developer did not receive a response.
尽管开发者已撤下Artex,但该框架的多个版本已被复制,并继续被开发和使用,与其他类似工具一同存在。Valadon表示,GitGuardian迄今已记录了45种不同的AI攻击框架,其中CyberstrikeAI、Hexstrike、CAIRN和PentAGI被研究人员频繁提及。它们的工作方式相似,技术能力略有差异。对于防御此类工具发起的攻击的人来说,一个好消息是,目前这些工具很容易被发现。
Despite the developer pulling Artex, multiple versions of the framework have been copied and continue to be developed and used, alongside other similar tools. Valadon says GitGuardian has documented 45 different AI attack frameworks so far, with CyberstrikeAI, Hexstrike, CAIRN, and PentAGI being frequently mentioned by researchers. They all work in similar ways, with some slightly different technical capabilities. One silver lining for people defending against attacks using such tools is that they’re currently easy to spot.
“这些工具目前并不隐蔽,”互联网安全公司Censys的首席安全研究员Silas Cutler表示。许多攻击会在系统中制造大量噪音——在短时间内发起成千上万次探测。例如,代理式编程平台Dosu表示,Artex被用于在短短90分钟内向其系统发送80,000个请求。研究人员表示,基本安全基础——修补软件、不使用默认凭据或暴露密钥、监控异常网络活动——可以帮助防御许多使用Artex的人所采取的方法。
“These tools are not stealthy right now,” says Silas Cutler, principal security researcher at internet security firm Censys. Many of the attacks create a lot of noise in systems—making thousands of probes in quick succession. For instance, agentic coding platform Dosu said Artex was used to send 80,000 requests to its systems in just 90 minutes. Researchers say that basic security fundamentals—patching software, not using default credentials or exposing secrets, and monitoring for unusual network activity—can help defend against many of the approaches from those using Artex.
要确定使用Artex和其他框架的网络犯罪分子数量并不容易;然而,根据Censys和Shadowserver基金会的分析,数千个此类系统被搭建并暴露在互联网上。也许暗示了使用Artex者的复杂程度,研究人员发现多个该工具未受保护且完全暴露在互联网上的实例,使他们能够查看操作者如何使用它。
Pinpointing the number of cybercriminals using Artex and other frameworks isn’t easy; however, thousands of instances of the systems being set up have been exposed to the internet, according to analysis by Censys and the Shadowserver Foundation. Perhaps hinting at the sophistication of those using Artex, researchers have found multiple instances of the tool left unsecured and fully open to the internet, allowing them to inspect how operators are using it.
Cutler表示,上周,两个暴露的Artex服务器似乎显示中文母语者正在针对其他中文实体。“在其中一个案例中,一所大学正遭受密码喷洒攻击,而且这一切都有学生试图修改成绩的感觉,”他说。
Last week, Cutler says, two exposed Artex servers appeared to show Chinese native speakers targeting other Chinese entities. “In one of the cases, it was a university that was being targeted with a password spraying attack, and it has all of that vibe of a student trying to change their grades,” he says.
“这种活动中存在一种层级,很可能是有组织的、复杂的犯罪活动,”卡特勒说。“与此同时,很可能还有一部分活动更多属于青少年犯罪范畴。”
“There is a level of this activity that’s occurring that is likely organized, sophisticated, criminal activity,” Cutler says. “There is also, in parallel, likely a subset of the activity that falls more into that juvenile crime realm.”