11-19 更新 实时
首页/ AI 大模型/ 正文

企业AI访问权限被贱卖,当心LLM混售商

AI 大模型 来源:ZDNet 2026-11-16 22:00 0 阅读 机器翻译
企业AI访问权限被贱卖,当心LLM混售商

Just_Super / E+ via Getty Images

Google 最新的 AI 威胁报告称,犯罪分子现在通过“LLM Mixers”转售窃取的企业 AI 访问权限。

AI Token 批次可以低至 $0.10 出售,而组织则要承担超过 $100,000 的使用账单。

支出上限、可疑行为警报和保持警惕可以防止你成为受害者。

短短几年内,企业已经采用 AI 和大型语言模型(LLMs)来管理从日历邀请到数据分析和编程的一切。恶意行为者通过 LLMjacking 窃取这些模型的凭证——而市场正在演变。

Google 关于 AI 威胁的最新报告,即来自 Google 威胁情报小组(GTIG)的 Q4 AI Threat Tracker,发现被称为 LLM Mixers 的网关正在以极低的价格出售访问权限。

LLM Mixer 的地下经济

LLMjacking 是窃取 AI 相关凭证的行为,包括用于访问 AI 模型的用户名和密码或 API 密钥。这使网络犯罪分子有机会使用属于受害组织(或个人)的 Token 和 AI 资源,这些资源通常采用按使用量计费或没有固定上限的订阅。

凭借这些凭证,未授权用户可以运行高资源消耗的计算任务、管理自己的恶意 AI 部署、从 AI 的记忆中提取并窃取企业数据,甚至投毒训练数据。更糟的是,如果使用需求超出受害者订阅的额度上限,网络犯罪分子还能推高用量成本,而最终为此买单的却是受害者。

GTIG 首席分析师 John Hultquist 此前表示,该网络安全部门在 2026 年观察到 LLMjacking 出现“大幅增长”,因此可以推断这一趋势很可能延续到 2027 年及以后。

AI 访问权限本身如今已成为一门利润丰厚的非法生意,由此催生了 Google 所称的 LLM Mixers:提供对 AI 服务及底层代理基础设施未授权访问的“非法店铺”。

你的 AI 凭证被出售的真实代价

在分析了俄语和中文交易市场上的 LLMJacking 供应商广告后,GTIG 研究人员发现,大多数供应商不仅提供对 AI 实验室前沿模型的访问,还提供被宣传为适合恶意用途的越狱模型。

然而,这些网络犯罪团伙似乎很难就“合理”定价达成一致,价格会因所提供的模型和可获取的凭证而剧烈波动。

如你所料,免费试用凭证的售价远低于企业凭证。商业用途的 Token 价格在 $0.10 到 $25 之间,其中一家供应商 APIMart 在提示词超过 100,000 Token 后,将 Token 价格提高了五倍。其他供应商则提供基于订阅的定价。

这在较低层级可能看起来不算什么,但Sysdig根据早期模型估算,LLMJacking每天可能造成约46,000美元的损失,而在最前沿的模型上甚至可能超过100,000美元。本月早些时候发布的Mandiant研究记录了多起事件,其中非法AI使用在短短24小时内就超过了100,000美元,部分组织的累计费用在访问权限被撤销前达到了100万美元。

非法LLM供应链

虽然直接凭证可以被出售,但基础设施元素也是一种有趣的服务 offering。Google 的威胁团队在报告中表示,LLM 基础设施的销售“将买家的流量通过一个被盗凭证和计算实例池进行路由,以确保可靠访问,同时还提供将查询量分散到多个受害者环境中的能力,以降低被检测到的可能性。”

LLM Mixer 看起来与网络犯罪地下经济的其他部分相似,在这个经济中,交易者往往专注于攻击链的一两个环节。例如,一个 LLM Mixer 提供利用企业 AI 模型所需的访问权限和基础设施,而它的合作伙伴可能会入侵受害者网络以获取初始访问权限,或从信息窃取型恶意软件的操作者那里购买被盗凭证。

如何保护你的组织

避免 LLMjacking 以及避免你组织的 AI 被加入 Mixer 的菜单,其方式类似于你防止入侵 SaaS 应用和云环境的方式:管理访问权限、监控并施加限制。

使用量和支出是首先要解决的领域。通过对每项 AI 服务和账户可签发及使用的 Token 数量设置限制,并在使用量突然激增时发出警报,组织可以在账户被攻破时限制损害,并在巨额账单累积之前加以阻止。

各组织还应设法从一开始就避免密钥落入网络攻击者之手。Google 建议使用密钥管理器来安全存储并频繁轮换 API 密钥,同时开展审计,以发现悄无声息的探测、配置错误或影子账户。

此外,全面部署钓鱼防护和多因素认证(MFA)也是明智之举,这能从一开始就让窃取凭证或入侵账户与网络的行为变得更难实施。

Charlie Osborne 是一名网络安全记者和摄影师,在伦敦为 ZDNET 和 CNET 撰稿。PGP 密钥:AF40821B 查看完整简介

译文由英文原文机器翻译,术语按行业惯例统一,可在文末展开原文核对。

查看英文原文(Your enterprise AI access can be sold for pennies – beware ‘LLM Mixers’)

Just_Super / E+ via Getty Images

Google’s latest AI threat report says criminals now resell stolen enterprise AI access through “LLM Mixers.”

AI token batches can be sold for as little as $0.10, while organizations foot usage bills of over $100,000.

Spending caps, suspicious behavior alerts, and vigilance can stop you from becoming a victim.

In just a few years, businesses have adopted AI and large language models (LLMs) to manage everything from calendar invites to data analysis and programming. Bad actors steal credentials for those models through LLMjacking — and the marketplace is evolving.

Google’s latest report on AI threats, the Q4 AI Threat Tracker from the Google Threat Intelligence Group (GTIG), found that gateways called LLM Mixers are selling access for cents on the dollar.

The LLM Mixer underground economy

LLMjacking is the theft of AI-related credentials, including usernames and passwords, or API keys, used to access AI models. This gives cybercriminals the opportunity to use tokens and AI resources belonging to a victim organization (or individual), typically with usage-based billing or subscriptions with no fixed cap.

With these credentials, an unauthorized user could run high-resource computing tasks, manage their own malicious AI deployments, extract and steal corporate data from the AI’s memory, or even poison training data. To add insult to injury, if demand outstrips the limits of the victim’s subscription, cybercriminals can rack up high usage costs that the victim is then on the hook for.

John Hultquist, chief analyst for GTIG, previously said the cybersecurity unit has seen a “major increase” in LLMjacking over 2026, so we can assume this trend will likely continue into 2027 and beyond.

AI access is now a lucrative illegal business in its own right, leading to what Google calls LLM Mixers: “illicit storefronts” that provide unauthorized access to AI services and underlying proxy infrastructure.

The true cost of your AI credentials being sold

After analyzing LLMJacking vendor advertisements on Russian- and Chinese-language marketplaces, GTIG researchers found that most offer access not only to frontier models from AI labs, but also to jailbroken models marketed as suitable for malicious purposes.

However, it doesn’t seem that many of these cybercriminal groups can agree on “reasonable” pricing, with prices swinging wildly depending on the models offered and the credentials available.

Free trial credentials fetch far less than enterprise credentials, as you would expect. Token prices for business use ranged from $0.10 to $25, and one vendor, APIMart, increased its token price fivefold once a prompt exceeds 100,000 tokens. Other vendors offer subscription-based pricing.

This might not seem like much at the lower tiers, but Sysdig estimates LLMJacking could cost around $46,000 based on earlier models, and even over $100,000 per day on cutting-edge models. Mandiant research published earlier this month recorded incidents with illicit AI usage exceeding $100,000 in only 24 hours, with cumulative charges at some organizations reaching $1 million before access was revoked.

The illicit LLM supply chain

While direct credentials can be sold, the infrastructure element is also an interesting service offering. Google’s threat team said in the report that the sale of LLM infrastructure “routes buyer traffic through a pool of stolen credentials and compute instances to ensure reliable access, while also offering the ability to spread query volume across multiple victim environments to reduce the likelihood of detection.”

LLM Mixers look similar to the rest of the cybercriminal underground economy, in which traders tend to specialize in one or two parts of an attack chain. For example, an LLM Mixer offers the access and infrastructure needed to exploit an enterprise AI model, while its partners might break into victim networks to secure initial access or buy stolen credentials from operators of infostealer malware.

How to protect your organization

Avoid LLMjacking and having your organization’s AI added to a Mixer’s menu similarly to how you would prevent intrusions into SaaS applications and cloud setups: managing access, monitoring, and imposing restrictions.

Usage and spending are the first areas to tackle. By setting limits on how many tokens can be issued and used across every AI service and account, and by alerting when usage suddenly spikes, organizations can limit damage if an account is compromised and stop massive bills before they rack up.

Organizations should also try to prevent their keys from ending up in the hands of cyberattackers in the first place. Google recommends using a secrets manager that securely stores and frequently rotates API keys, and running audits to detect quiet probing, misconfigurations, or shadow accounts.

It’s also advisable, across the board, to implement phishing protections and multi-factor authentication (MFA) to make the business of stealing credentials or compromising your accounts and networks more difficult in the first place.

Please accept cookies to view the comments powered by Disqus.

Charlie Osborne is a cybersecurity journalist and photographer who writes for ZDNET and CNET from London. PGP Key: AF40821B See full bio

原文为英文,正文由自动抓取并保留原文。

本文由采集器自动同步

原文发布于 ZDNet(Your enterprise AI access can be sold for pennies – beware ‘LLM Mixers’),标题、摘要与正文由机器翻译,术语按行业惯例统一,未经人工逐句校订。

阅读原文 ↗