Data Governance · 3 min read
LLM Mixers sell AI access from $0.10 as account bills reach $1M
Google’s Q4 AI Threat Tracker finds stolen enterprise AI access sold through LLM Mixers, with business tokens priced at $0.10 to $25 and illicit use exceeding $100,000 in 24 hours.
Key takeaways
- Q4 AI Threat Tracker: LLM Mixers resell unauthorized access to frontier and jailbroken AI models.
- Business-use token prices ranged from $0.10 to $25 in Russian- and Chinese-language marketplace advertisements.
- APIMart increased its token price fivefold once a prompt exceeded 100,000 tokens.
- Sysdig estimated about $46,000 in costs on earlier models and more than $100,000 per day on cutting-edge models.
- Mandiant documented more than $100,000 in illicit AI usage within 24 hours; some organizations reached $1 million in cumulative charges.
Marcus Bennett
Stolen enterprise AI credentials can now fetch $0.10 per business token, while Mandiant has documented more than $100,000 in illicit AI usage within 24 hours and cumulative charges reaching $1 million at some organizations.
Google’s Q4 AI Threat Tracker, produced by the Google Threat Intelligence Group (GTIG), identifies a resale market for unauthorized AI access that Google labels LLM Mixers. Within a few years, enterprises adopted AI and large language models (LLMs) for calendar invitations, data analysis and programming. LLM Mixers turn that adoption into a source of billable attack infrastructure.
What is an LLM Mixer?
Google defines LLM Mixers as illicit storefronts selling unauthorized access to AI services and their underlying proxy infrastructure, the routing layer for model traffic. LLMjacking is credential theft: criminals obtain usernames, passwords or application programming interface (API) keys used to reach AI models. Stolen access usually carries usage-based billing or a subscription without a fixed cap.
With those credentials, criminals can run compute-heavy jobs, operate malicious AI deployments, extract and steal corporate data held in AI memory, or poison training data. When demand exceeds the victim’s limits, attackers keep using the account and the organization absorbs the charge.
GTIG chief analyst John Hultquist said his unit had seen a “major increase” in LLMjacking over 2026, a trajectory likely to persist into 2027 and beyond.
How cheap is illicit access?
GTIG researchers analyzed LLMjacking advertisements on Russian- and Chinese-language marketplaces. Most sellers offered access to frontier models (leading systems from AI laboratories) and jailbroken models marketed as suitable for malicious purposes. Pricing lacked consensus and swung sharply with model choice and credential quality.
Free trial credentials fetched less than enterprise credentials. Business-use prices for tokens, the billable units used to process prompts, ranged from $0.10 to $25. APIMart increased its token price fivefold once a prompt exceeded 100,000 tokens, while other vendors used subscription pricing.
At one end, a token batch costs $0.10; at the other, some compromised organizations have accumulated $1 million before access was revoked. Infrastructure can distribute query volume across several victim environments, so the entry price does not cap aggregate loss.
Sysdig estimates potential LLMjacking costs at around $46,000 based on earlier models and more than $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. Cumulative charges reached $1 million at some organizations before revocation.
How does the illicit supply chain work?
GTIG said the infrastructure trade “routes buyer traffic through a pool of stolen credentials and compute instances to ensure reliable access.” The service also spreads query volume across multiple victim environments, reducing the likelihood of detection. This separates access resale from the intrusion work required to obtain credentials.
The market mirrors the wider cybercriminal economy’s specialization. An LLM Mixer supplies AI access and proxy infrastructure, while partners may break into victim networks for initial access or buy stolen credentials from operators of infostealer malware. Each participant can focus on one or two stages of the attack chain.
What should data teams do now?
GTIG frames the defense like software-as-a-service (SaaS) and cloud intrusion prevention: manage access, monitor behavior and impose restrictions. For data teams, those controls need to cover every AI service and account.
- Set issuance and consumption limits for tokens on every AI service and account.
- Alert on sudden usage spikes so teams can stop large bills before they accumulate.
- Store API keys in a secrets manager and rotate them frequently.
- Audit for quiet probing, misconfigurations and shadow accounts.
- Apply phishing protections and multi-factor authentication (MFA) across the organization.
Together, the controls address both sides of the risk: they make stolen credentials harder to obtain and limit damage when an account still fails. For data leaders, the operating test is whether token issuance, consumption and anomalies remain visible enough to cap the bill.
cloud.google.com (Original)
More from Marcus Bennett
Show full bio
News editor covering people, launches and day-to-day practice in Analytics Digest is a professional daily at Analytics Digest.
10 articles