LLMjacking 攻擊竊取企業 AI 存取憑證,每日帳單可超過 $100,000
LLMjacking Attacks Steal Corporate AI Access, Run Up $100,000 Daily Bills
LLMjacking 攻擊者竊取企業 LLM 服務的 API 金鑰及認證 Token,冒充合法用戶調用模型,相關費用由受害企業承擔。Mandiant 指每日詐騙費用可超過 $100,000,個別企業在撤銷攻擊者存取權前累計費用達 $1 million。Google 建議集中管理機密憑證、定期輪替金鑰、落實最小權限,並自動掃描程式碼中的憑證。

Cybercriminals are increasingly stealing corporate access credentials for large language model (LLM) services to exploit expensive AI computing resources and services without authorization. Known as “LLMjacking,” the attacks involve stealing corporate LLM application programming interface (API) keys and authentication tokens to run their own AI services or conduct large-scale model inference. The resulting bills, however, are charged to the victim companies.
According to a report released November 2 by Mandiant, Google’s threat intelligence unit, fraudulent charges resulting from unauthorized resource consumption can exceed $100,000 a day. In some cases, cumulative charges at affected companies reached $1 million before attackers’ access was revoked.
The primary targets in LLMjacking attacks are API keys and access tokens, which serve as proof of authorization to use an API.
An API key is an authentication credential that an application or service uses to verify its authorization when sending requests to an external LLM. Once attackers obtain the key, they can access cloud-based AI services and invoke models while posing as legitimate users. They can exploit the privileges associated with previously issued credentials without obtaining account passwords or going through normal login procedures.
LLMjacking attacks have increased steadily over the past two years. CrowdStrike's August report, 2026 Threat Hunting Report, documented an extreme case in which an attacker using stolen tokens generated about 200,000 API requests in just two minutes.
An attack group tracked by Google as UNC6993 illicitly obtained LLM API keys, used them to generate additional keys and activated services within the victim company's cloud environment. The resulting charges reached $100,000.
In another case, an attacker enrolled a victim company's account in a Pay-As-You-Go plan to gain unauthorized access to premium AI models.
According to The Register, a particularly notable aspect of this attack was the use of a “burn and rotate” strategy. Attackers sent large volumes of requests in a short period. When the company’s predefined usage quotas or rate limits were reached, they deleted existing service accounts and created new ones to circumvent the restrictions.
◇Attackers Use Stolen API Keys to Operate Low-Cost AI Services
The attacks have also allowed cybercriminals to generate additional revenue. In another intrusion, Google found that an attacker had stolen a third party's API keys and used them to operate an LLM routing service called Jet Night. The attacker made more than 500,000 LLM API calls in less than eight hours.
LLM routing services, commonly offered through AI gateway platforms, act as intermediaries that direct users' AI requests to multiple LLM providers and models. Such services generally operate using their own API keys or through formal agreements with model providers.
Because the operator did not pay the legitimate API usage charges and instead shifted the costs to other companies' accounts, it was able to offer services at prices far below market rates.
Jet Night advertised access to cutting-edge models and offered an unlimited token plan for about $60 a month. According to Google, comparable services would typically cost consumers between $100 and $250 a month.
◇LLMjacking Emerges as a More Attractive Target Than Cryptojacking
Google expects cybercriminals to increasingly prioritize LLMjacking over cryptojacking. Cryptojacking involves stealing cloud computing resources to mine cryptocurrencies. Attackers typically install malicious code on other people's computers or cloud servers to use their central processing units (CPUs) and graphics processing units (GPUs) without authorization.
The spread of agentic AI is another factor increasing the threat posed by LLMjacking. When attackers combine stolen LLM access credentials with agentic tools, they can sustain large volumes of API calls and automated attacks. This allows them to scale up the size and speed of their operations with relatively little effort.
Google recommends that companies prioritize strengthening identity and access management (IAM) for cloud and AI services, as well as software supply chain security.
According to Google, several measures are essential:
Introduce centralized secrets management platforms to securely store and manage critical credentials, including API keys and authentication information.
Implement regular key rotation to prevent attackers from using compromised credentials over extended periods.
Apply the principle of least privilege, granting users and service accounts only the access rights necessary to perform their assigned tasks.
The principle of least privilege means, for example, that if a particular service account only needs to call a single LLM API, it should not also be granted access to other cloud projects or additional models. This limits the resources an attacker can reach even if the account is compromised.
Companies should also continuously inspect source code for embedded API keys and authentication credentials. Developers sometimes hard-code API keys or passwords into source code for convenience. If a code repository is subsequently exposed, those credentials can be leaked along with it. Google recommends automating secrets scanning across codebases to identify and remove API keys and other sensitive authentication information.
來源:The Elec · thelec.net