Safety & Security

Stolen Enterprise AI Access Sells for Ten Cents on Underground 'LLM Mixers'

Google's Q4 AI Threat Tracker details 'LLM Mixers': illicit storefronts reselling stolen enterprise AI access from $0.10 per token, with victim bills reaching $1 million.

Your enterprise AI access can be sold for pennies – beware ‘LLM Mixers’
Your enterprise AI access can be sold for pennies – beware ‘LLM Mixers’cliff1126 / pixabay
By Sophie Lindqvist4 min read

Why it matters

  • Google's Q4 AI Threat Tracker found 'LLM Mixers' reselling stolen enterprise AI token batches for as little as $0.10.
  • Sysdig estimates LLMjacking costs around $46,000 on earlier models and over $100,000 per day on cutting-edge models.
  • Mandiant recorded illicit AI usage exceeding $100,000 in 24 hours, with cumulative charges reaching $1 million at some organizations.
  • Vendor APIMart raises token prices fivefold for prompts exceeding 100,000 tokens.
  • GTIG's John Hultquist reported a 'major increase' in LLMjacking over 2026.

Stolen access to enterprise AI models is being resold for as little as $0.10 per token batch through underground gateways Google calls "LLM Mixers" — while victim organizations pick up bills that can exceed $100,000 in a single day. That finding comes from Google's Q4 AI Threat Tracker, published by the Google Threat Intelligence Group (GTIG), and it marks a shift: stolen AI access has become a standalone illegal business rather than a side effect of ordinary account theft.

The economics are brutal for anyone on the wrong end. Cybercriminals steal AI credentials — usernames, passwords, or API keys — in attacks Google labels LLMjacking. Because most enterprise AI services use usage-based billing or subscriptions without a fixed cap, an intruder can run compute-heavy jobs, deploy their own malicious AI services, extract corporate data from a model's memory, or poison training data. The victim pays for all of it.

What exactly is an LLM Mixer?

Google describes LLM Mixers as "illicit storefronts" that sell unauthorized access to AI services along with the proxy infrastructure behind them. The name explains the mechanism: buyer traffic gets routed through a pool of stolen credentials and compute instances, which keeps access reliable and spreads query volume across multiple victim environments so no single account shows an obvious spike.

That last part is the clever bit. By distributing queries across many victims, Mixers reduce the likelihood that any one organization notices anomalous usage. It is the same specialization logic that runs the rest of the cybercriminal underground, where traders focus on one or two links in the attack chain. An LLM Mixer sells access and infrastructure; its partners break into networks for initial access or buy credentials harvested by infostealer malware operators.

The takeaway: this is an organized supply chain, not opportunistic freeloaders piggybacking on a leaked password.

How much does stolen AI access cost?

GTIG researchers analyzed LLMjacking vendor advertisements on Russian- and Chinese-language marketplaces. Most vendors offer access not only to frontier models from major AI labs but also to jailbroken models explicitly marketed as suitable for malicious purposes.

Pricing is anything but standardized — it swings wildly depending on the models and credentials on offer:

  • Free trial credentials sell for far less than enterprise credentials.
  • Token prices for business use range from $0.10 to $25.
  • One vendor, APIMart, raises its token price fivefold once a prompt exceeds 100,000 tokens.
  • Other vendors sell subscription-based plans instead of per-token batches.

Those price tags look small, but the damage on the victim's side is not. Sysdig has previously estimated that LLMjacking can cost roughly $46,000 based on earlier models, and over $100,000 per day on cutting-edge ones. Mandiant research published earlier this month recorded incidents where illicit AI usage exceeded $100,000 within 24 hours, and where cumulative charges at some organizations reached $1 million before access was finally revoked.

The gap between a dime-a-token resale and a million-dollar corporate bill is the whole business model.

Is LLMjacking getting worse?

Yes, by Google's own account. John Hultquist, chief analyst for GTIG, has said the cybersecurity unit saw a "major increase" in LLMjacking over 2026 — and the reasonable assumption, on that trajectory, is that the trend continues into 2027 and beyond.

The context matters here. In just a few years, businesses have adopted AI and large language models for everything from calendar invites to data analysis and programming. Every new enterprise deployment is another set of credentials that can be lifted, pooled, and resold. The attack surface grew faster than the guardrails around it.

The takeaway: the more a company wires AI into daily operations, the more it has riding on credential hygiene.

How can organizations protect themselves?

Google's guidance treats AI services like any other SaaS or cloud asset: manage access, monitor behavior, and impose restrictions. The first target is usage and spending. Organizations should set limits on how many tokens can be issued and used across every AI service and account, and configure alerts for sudden usage spikes. A hard cap turns a catastrophic bill into a bounded one.

The second target is keeping credentials out of criminal hands in the first place:

  • Use a secrets manager that securely stores and frequently rotates API keys.
  • Run regular audits to catch quiet probing, misconfigurations, and shadow accounts.
  • Implement phishing protections and multi-factor authentication (MFA) across the board to make credential theft harder and stolen passwords less useful.

None of this is exotic. It is standard SaaS security discipline applied to a new class of asset — which is precisely the point. The organizations losing $1 million to LLMjacking were not missing futuristic defenses; they were missing caps, rotation, and monitoring.

The takeaway: the cheapest defense is a spending limit, and most victims simply never set one.

What happens next?

The Mixer economy will likely keep maturing as long as enterprise AI spend keeps climbing and uncapped billing stays common. Google's report effectively tells security teams to assume their AI credentials already have street value — ten cents at a time — and to budget accordingly.

Original: cloud.google.com

Share this article:

More from Sophie Lindqvist

Sophie Lindqvist

Show full bio

Staff writer covering regulation, policy and their impact on artificial intelligence and machine learning at AI In Context.

749 articles

Related articles

  1. Stolen AI Credentials Fuel a Booming Black Market in LLMjacking
  2. Google Warns LLMjacking Attacks Can Burn $100,000 in Cloud Charges in a Day
  3. Microsoft Disrupts EvilTokens, AI Chatbot Platform Behind 12,000 Account Hacks
  4. Google Cloud warns attackers will turn from data theft to AI decision manipulation
  5. OpenAI Bans Accounts Linked to DPRK Threat Actors Using AI for Intrusion Research

« Previous articleNext article »