Your enterprise AI access can be sold for pennies – beware ‘LLM Mixers’

Google's Threat Intelligence Group (GTIG) has published its Q4 AI Threat Tracker, which documents an evolving marketplace around stolen access to enterprise AI services. The report introduces the term "LLM Mixers" for illicit storefronts that sell unauthorized access to AI services along with the proxy infrastructure behind them. The practice at the center of it, LLMjacking, is the theft of AI-related credentials — usernames, passwords or API keys — that grant access to large language models. Because most commercial AI services bill per token or through subscriptions without a hard ceiling, whoever holds a stolen key can consume resources that the legitimate account owner pays for.
The economics are stark. GTIG researchers analyzed advertisements from LLMJacking vendors on Russian- and Chinese-language marketplaces and found that most offer access to frontier models from major AI labs as well as jailbroken models marketed for malicious use. Pricing varies widely by model and credential type. Token prices for business-oriented access ranged from $0.10 to $25, and one vendor, APIMart, reportedly multiplied its token price fivefold once a prompt passes 100,000 tokens. Some sellers instead offer subscription-style pricing. Free trial credentials sell for far less than enterprise ones.
The downstream bills dwarf the resale prices. Citing Sysdig estimates, the report puts potential LLMjacking costs at roughly $46,000 under earlier models and over $100,000 per day on cutting-edge models. Mandiant research published earlier in the month recorded incidents in which illicit AI usage exceeded $100,000 within 24 hours, with cumulative charges at some organizations reaching $1 million before access was revoked.
What distinguishes the mixer model from simple credential resale is the infrastructure layer. Google's threat team describes LLM infrastructure offerings that route buyer traffic through a pool of stolen credentials and compute instances to keep access reliable. The same routing lets a seller spread query volume across multiple victim environments, which reduces the chance that any single organization notices the anomaly. This suggests the mixer is essentially a load-balancing layer built on compromised accounts — an arrangement that makes detection harder precisely because no one victim sees the full traffic pattern.
The mixers fit the division-of-labor pattern common in cybercriminal markets, where participants specialize in one or two links of an attack chain. A mixer supplies the access and routing needed to exploit an enterprise AI model, while partners handle initial intrusion into victim networks or supply credentials harvested by infostealer malware. John Hultquist, chief analyst for GTIG, previously described a major increase in LLMjacking over 2026, which points to the trend continuing.
Once inside, an unauthorized user can do more than run up a bill. The report notes they could run high-resource computing tasks, operate their own malicious AI deployments, extract corporate data held in a model's memory, or poison training data. Each of those is a different risk profile: cost overrun is the most visible, data exposure is the most damaging, and training-data poisoning is the hardest to detect after the fact.
The defensive advice mirrors cloud and SaaS security practice. The first lever is usage and spending control: cap the number of tokens issued and consumed across every AI service and account, and alert on sudden spikes so a compromised key is contained before a large bill accumulates. The second is key hygiene — a secrets manager that stores API keys securely and rotates them frequently, plus audits that surface quiet probing, misconfigurations or shadow accounts. Third is reducing the chance credentials are stolen at all, through phishing protections and multi-factor authentication. A broader point, which is my own reading rather than something the report states, is that per-token billing turns every leaked key into an open-ended financial liability in a way that seat-based SaaS licensing historically did not, since a stolen seat is one seat and a stolen key can be consumed without limit.
Separately, the report cites Ramp AI index data on corporate AI spending: among firms paying for AI, the top 1% spent over $10,000 per employee per month, compared with about $16 for the median firm. That skew matters for exposure. Organizations with heavy AI consumption have the most to lose from a hijacked account, while smaller outfits that spend little may be the least likely to have set spending caps in place — a gap the report explicitly flags as something smaller companies tend to overlook.
Why it matters: Any team issuing long-lived API keys for LLM providers is a potential supplier to this market, whether or not it has noticed. Developers and platform teams should treat model credentials with the same rigor as cloud root keys — scoped permissions, rotation, per-key budget alarms — and finance and security staff should agree on hard spend ceilings per service, since inference billing can scale from routine to six figures within a day. The mixer routing detail is the part worth internalizing: because traffic is spread across victims, per-account volume thresholds alone may not catch an intrusion, so anomalous usage patterns and key provenance likely need attention as well.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
Underground brokers are selling access to hijacked enterprise AI accounts, with companies footing the usage bill, Google reports.