LLMjacking Turns Stolen Enterprise AI Access Into a Costly Underground Service

| 2 Min Read
Criminals are reselling stolen cloud and model credentials, driving AI charges above $100,000 a day while probing for ways to stay hidden.

Stolen AI credentials have become a subscription business: victims pay the cloud bill while criminals sell access to their models. Google Threat Intelligence Group (GTIG) says some LLMjacking campaigns have generated more than $100,000 in charges in a day; in several cases, the total reached $1 million before access was revoked. The sums show how quickly abuse can scale when a credential grants access to paid model services and requests can be routed through it in bulk.

GTIG’s findings draw on incidents from 2026 involving cloud and large language model providers across North America, South America, and Asia Pacific. The pattern points to a costly blind spot: credentials issued for one purpose can be reused to serve customers the victim never authorized. A cloud account may therefore look like it’s operating normally even as outsiders consume its model capacity. That makes the cost harder to distinguish from legitimate use until an alert, an investigation, or the bill brings it into view.

Stolen credentials fuel AI resale services

GTIG describes the resale infrastructure as “LLM Mixers”: criminal gateways that combine stolen API keys, route customer requests through compromised accounts, and spread usage across multiple credentials. That distribution can obscure where requests originate and make a single compromised account harder to spot. It also turns stolen access into a service: buyers can send prompts to a gateway without handling the underlying keys themselves. (And this is the part most people overlook.) The criminal operator doesn’t need to break into each customer’s systems; the victim’s credentials do the work.

One operation, tracked as UNC6993, used open-source tools including new-api and gcli2api to pool stolen keys and route requests across accounts. In another case, stolen third-party keys powered Jet Night, a token-routing service. GTIG reported that it handled more than 500,000 LLM API calls in under eight hours, passing nearly $1 million in token costs to the victim. Those details describe two specific cases, but the underlying model is straightforward: a gateway can hide the relationship between the customer making a request and the account paying for it.

Jet Night advertised unlimited access to frontier models for about $60 per month, far below the usual price of comparable legitimate access. GTIG suspects occasional outages followed victims discovering the compromise and revoking keys. The low advertised price shifts the cost elsewhere: the seller collects subscription revenue while the account holder absorbs the usage charges. This is more significant than it looks. A service that appears cheap to its customers can create a substantial and sudden expense for the organization whose access has been stolen.

“LLM Mixers turn stolen enterprise model access into a service for paying customers, while the victim’s systems continue accepting requests through credentials they recognize,” said Piyush Sharma, co-founder and CEO at Tuskira. He said credentials should be limited to the applications, models, and workloads that need them, rather than serving as reusable access across an organization. That advice addresses a basic weakness in credential security: permissions that are broader or longer-lived than a task requires leave more room for misuse if a key is exposed.

Quiet tests can precede a spending spike

The bill may arrive long after the first warning signs. GTIG says attackers gain entry through exposed or stolen cloud identity tokens, including IAM credentials, service account keys, GitHub personal access tokens, and active session tokens. They also look for API keys in public repositories and exposed model endpoints. These routes differ, but they can lead to the same practical outcome: a valid credential that lets an attacker make requests under someone else’s account.

Before putting stolen keys to heavy use, attackers may test them with tiny requests, sometimes setting max-tokens=1. The resulting usage can fall beneath conventional spending alerts. “Testing a stolen key with a single-token request barely registers as consumption,” said Nick Tausek, lead security automation architect at Swimlane. He said a small request can warrant investigation when it coincides with a new credential or an unexpected permission change. The signal is contextual: a low bill by itself may look harmless, while the same request alongside an unfamiliar identity change deserves a closer look.

After gaining a foothold, attackers may search cloud services, repositories, and data stores for more credentials. GTIG observed tactics including privilege escalation, unauthorized service accounts, injected SSH keys, and persistence through compromised workloads. Each step can make the original compromise harder to contain, because revoking one exposed key may not remove access created elsewhere. For security teams, tracing related identity and permission changes can matter as much as locating the first credential.

In one follow-on incident, an attacker exploited a user ID rebinding vulnerability in the LiteLLM gateway to turn ordinary virtual keys into administrative credentials. Sharma said compromising the routing layer can extend access beyond the key that was first stolen. That raises the stakes for gateway security: a routing component can sit between users and model providers, so weaknesses there may affect the reach of credentials passing through it.

Contain access before usage scales

GTIG recommends stronger credential management, least-privilege permissions, narrowly scoped API keys, alerts for suspicious low-volume requests, and spending limits for model usage. It also advises monitoring for unauthorized AI service enablement and gateway vulnerabilities. These measures address different parts of the chain: limiting what a stolen key can do, spotting early signs of testing, and reducing the time or cost available to an attacker. No single alert is likely to tell the whole story.

The operational takeaway is to treat AI credentials as production access: scope them to specific models and workloads, then connect small anomalous requests with changes to identities and permissions. If you're working in this space, that means looking beyond large usage spikes and asking whether a seemingly minor request came from an expected credential, application, and workload. A small test can be easy to miss. But when teams can connect it to a new identity or permission change, they may have a chance to cut off resale before a handful of quiet tests turns into hundreds of thousands of calls.

What the pattern means for AI services

LLM Mixers expose a gap between controlling who can sign in and controlling what a valid credential can spend. In similar cloud services, a credential can remain technically legitimate even when its use has shifted to an unauthorized purpose. That makes monitoring model access a cost and security problem at once: usage records, identity changes, and permissions need to be understood together.

The incidents also show why access design matters as AI services become part of ordinary business workflows. Broad credentials make setup easier, but they can also give a stolen key more reach than the application that originally needed it. Narrow scopes and spending limits can constrain that exposure, while alerts tuned to quiet activity may help teams catch probing before it becomes a costly stream of requests. The evidence GTIG describes doesn’t establish how common this resale model is across the industry, but it does show that the method can turn stolen access into a commercial service—and leave the victim paying for it.

Source: David Smith · www.csoonline.com

Comments

Please sign in to comment.