LLMjacking: Victims Charged Up to $1 Million

Stolen access to artificial intelligence services is fueling a black market that Google is monitoring. For victimized companies, the risks also involve data and the malicious use of their resources.
Mandiant Reports Costs Reaching $1 Million
Mandiant reported in research published earlier this month that illicit AI usage costs exceeded $100,000 within 24 hours. Some organizations had accumulated costs of up to $1 million by the time access was revoked. Sysdig estimates the potential cost at around $46,000 with older models, and over $100,000 per day with cutting-edge models. When fraudulent usage exceeds the limits of a subscription, the additional bill may fall on the victim.
Controlling access, monitoring it, and enforcing restrictions are the recommended protections, similar to those for SaaS or cloud services. Capping the issuance and consumption of tokens for each account and service can limit losses after a compromise. Alerts triggered by a sudden spike in usage can also help prevent costs from accumulating.
Google recommends storing API keys in a secret manager and renewing them frequently. The company also advocates for audits that look for discreet reconnaissance attempts, misconfigurations, or ghost accounts. Protections against phishing and multi-factor authentication are also part of the recommended measures to make it more difficult to compromise credentials, accounts, and networks.
Mixers Distribute Requests Among Multiple Victims
Google refers to "LLM mixers" as illegal businesses that resell access to AI services using proxy infrastructure. According to its threat team, customer requests pass through a set of stolen credentials and computing instances to maintain reliable access. Their distribution across multiple compromised environments aims to make detection less likely.
Le brief IA que les pros lisent chaque soir
Les 7 actus IA du jour, décryptées en 5 min. Gratuit.
Inclus dès l'inscription : notre sélection des meilleurs guides & comparatifs IA.
Choisissez votre rythme
Gratuit · Pas de spam · Désabonnement en 1 clic
These intermediaries provide access to the enterprise model as well as the technical means to exploit it. In this underground economy, sellers typically focus on one or two components of the attack. Their partners may handle the initial intrusion into a network or acquire credentials from infostealer operators, information-stealing malware.
LLMjacking precisely begins with the theft of authentication means: username, password, or API key. These grant access to tokens and resources of a company or individual, usually billed on a consumption basis or associated with a subscription that lacks a fixed cap. An intruder can then launch resource-intensive computations or exploit their own malicious deployments. Possibilities also include stealing company information stored in the AI's memory and poisoning the data used for training.
APIMart Adjusts Pricing Based on Prompt Size
The GTIG examined listings on Russian and Chinese-speaking markets. According to its researchers, the majority of sellers studied market both access to cutting-edge models from labs and unrestricted models, presented as usable for malicious purposes. Prices vary significantly depending on the available model and the credentials offered, with trial account credentials costing significantly less than those from enterprises.
For tokens intended for commercial use, the GTIG notes a range from $0.10 to $25. At APIMart, the price of tokens increases fivefold once the prompt exceeds 100,000 tokens. Other sellers adopt a subscription model.
John Hultquist, chief analyst at GTIG, stated that his team observed a significant increase in LLMjacking in 2026. A continuation of this trend in 2027 and beyond seems likely, though it is not a guaranteed outcome.
Brief IA — L'actualité IA en français
L'essentiel de l'actualité de l'intelligence artificielle, décrypté et expliqué chaque jour.