Stolen Access to Corporate AI Sold From $0.10, and Victims Pay Up to $100,000 a Day
Hackers steal API keys for corporate AI models and resell them on underground storefronts that Google calls "LLM Mixers." A token pack, a unit the provider uses to bill for model work, costs from $0.10 to $25 there, and the bill goes to the key's owner. Mandiant has seen cases above $100,000 in 24 hours. It's as if a stranger ran heavy queries all night through a key you forgot in a repository. That's why you should turn on spending limits and spike alerts before a breach.
AI-processed from ZDNet AI; edited by Hamidun News
Imagine your developer accidentally leaves an AI model API key in a public repository. Until recently, the main risk was data leakage. Now there's a direct financial hit: hackers list that key on an underground "storefront," the buyer pays from $0.10 per token pack (a token is the unit a provider uses to measure how much work a model does), and you get the entire real bill for the usage. Google Threat Intelligence Group (GTIG) called these storefronts "LLM Mixers" in its Q4 AI Threat Tracker report. According to Mandiant researchers, illegal AI use at some companies exceeded $100,000 in just 24 hours.
What LLMjacking is and why "mixers" are involved
LLMjacking is the theft of credentials for accessing language models: logins, passwords, and API keys. If a company's bill depends on how much it uses or there's no hard ceiling on its subscription, an attacker gets someone else's tokens and computing power. They can run heavy workloads, deploy their own malicious AI services, extract corporate data from the model's "memory," or even poison training data.
Thieves used to use the key themselves. Now the theft has become a separate business. GTIG reviewed seller listings on Russian-language and Chinese-language forums. Most offer not only access to advanced models from major labs but also "jailbroken" versions, models with restrictions removed, advertised as suitable for malicious purposes.
A "mixer" is not just a key shop. According to Google's description, it routes a buyer's traffic through a pool of stolen credentials and compute instances. Requests are spread across several compromised environments to reduce the chance of detection. For each victim, the activity looks like minor odd behavior, while the buyer gets a steady channel in aggregate.
The scheme resembles the rest of the cybercrime market: sellers specialize in one or two links of the chain. The mixer provides access and infrastructure, while partners break into networks or buy stolen passwords from infostealer operators, malware that harvests credentials.
What it costs and who pays
Prices on the underground market swing widely. Trial accounts cost noticeably less than corporate ones. For business-grade access, token prices ranged from $0.10 to $25. One seller, APIMart, raises its price fivefold when a request exceeds 100,000 tokens. Some sellers also offer subscriptions.
For the buyer, this is pocket change. For the victim, the bill is something else entirely. Sysdig estimates LLMjacking damage at around $46,000 on early models and more than $100,000 a day on the most advanced ones. Mandiant research published in early November recorded cases where illegal AI use exceeded $100,000 in 24 hours. At some organizations, accumulated charges reached $1 million before access was revoked.
Ramp AI Index data provides context for these figures. Among companies that pay for AI, the top 1% spend more than $10,000 per employee per month, while the median company spends about $16. At a company with a large AI budget, stolen access can stay hidden for a long time among normal spending. Small firms often don't set limits at all.
How to protect yourself
Google recommends treating AI access the same way as SaaS applications and cloud infrastructure: manage access, monitor it, and limit it.
- Spending limits. Set token and dollar ceilings for each AI service and account. If a key is stolen, the damage stays contained.
- Spike alerts. A sharp rise in usage at night or on weekends is a reason to block the key immediately.
- Secrets manager. Store API keys in a dedicated vault and rotate them regularly.
- Audits. Look for quiet probing, configuration errors, and shadow accounts nobody remembers.
- MFA and phishing protection. They make credential theft harder at the first step.
What this means for developers and product managers
An AI key is now an asset with a monetary equivalent, much like a payment card. If you're building agents or internal assistants, check where the keys are stored, who has them, and whether a hard budget is set on the provider's side. Don't wait for the end-of-month invoice.
The report has limits: the prices come from underground listings and vary widely across models and access types, and the damage estimates are approximate. But the trend is clear. GTIG's chief analyst John Hultquist has said the number of LLMjacking cases grew noticeably in 2026, and it is unlikely to fall in the near future.
- Nov 21, 2026A swimming team captain handed ChatGPT three scheduled tasks and cut her routine work by a multiple
- Nov 21, 2026MIRA Neural Network Renders Rocket League at 20 FPS, but Ping Hits 460 ms
- Nov 21, 2026Google built a closed copy of the internet with fake banks to train Chrome's AI agents
- Nov 21, 2026MLPerf Training Adds Nvidia Rubin and TPU Ironwood in First Appearance: 153 Systems
Key Questions Answered
What LLMjacking is and why "mixers" are involved?
LLMjacking is the theft of credentials for accessing language models: logins, passwords, and API keys. If a company's bill depends on how much it uses or there's no hard ceiling on its subscription, an attacker gets someone else's tokens and computing power. They can run heavy workloads, deploy their own malicious AI services, extract corporate data from the model's "memory," or even poison training data. Thieves used to use the key themselves. Now the theft has become a separate business.
What it costs and who pays?
Prices on the underground market swing widely. Trial accounts cost noticeably less than corporate ones. For business-grade access, token prices ranged from $0.10 to $25. One seller, APIMart, raises its price fivefold when a request exceeds 100,000 tokens. Some sellers also offer subscriptions. For the buyer, this is pocket change. For the victim, the bill is something else entirely. Sysdig estimates LLMjacking damage at around $46,000 on early models and more than $100,000 a day on the most advanced ones.
How to protect yourself?
Google recommends treating AI access the same way as SaaS applications and cloud infrastructure: manage access, monitor it, and limit it. Spending limits. Set token and dollar ceilings for each AI service and account. If a key is stolen, the damage stays contained. Spike alerts. A sharp rise in usage at night or on weekends is a reason to block the key immediately. Secrets manager. Store API keys in a dedicated vault and rotate them regularly. Audits.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.