LLM Jacking: Stolen Keys, Resold Model Access

Mandiant saw stolen model access run up six-figure bills in a day. LLM jacking is credential theft pointed at your token quota, then resold.

By Muhammad Saif and Aashan Khan · Techtribe co-founders

LLM jacking is what it is called when someone steals the right to call your models and spends it until the invoice screams. Google's Threat Intelligence Group and Mandiant put the name on a pattern they had already been investigating, then restated it on November 16, 2026 inside a wider report on agentic threats. The longer note is the November 2 write-up by Christopher Romano, Codi Starks, and their colleagues. The pattern is not a clever prompt. It is credential theft pointed at a metered API, followed by a storefront that resells the meter.

The bills they are willing to put in public are the reason to care this week. In incidents through 2026, single-day fraudulent charges exceeded $100,000. In the worst cases they describe, the total reached $1 million before someone revoked access. Cryptojacking was a slow leak. This is a spike. If your alert for "AI spend" is a monthly review, you will meet the spike in the invoice, which is the remediation they say came too late.

A related theft, a vendor key copied out of a training bucket, is the story in the Fireworks breach and the story of what that key was used for in the Cursor extension incident. Those posts are about a specific vendor and a specific marketplace. This one is the criminal market those keys fall into, and the controls that make your key a bad product to resell.

How the access is taken

Mandiant's dominant initial-access story is not a jailbreak. It is a stolen identity. IAM credentials, service-account keys, GitHub personal access tokens, and live session tokens. They also describe keys left in public repositories, and a validation trick that should change how you alert: automated checks that send a one-token request, enough to see if the key works, small enough to hide under a volume alarm. If your detector only fires on a flood, the scout has already sorted the live keys from the dead ones.

They name other paths in specific intrusions, still as their observations. A misconfigured service-account key. A page that impersonates a portal and asks the user to paste a session payload, which skips the password and the second factor because the token already exists. Cloud federation tokens used to mint fresh keys and sidestep MFA. A stolen personal access token combined with a bug in a package running in a container. Public model endpoints, including local servers left open, tested by callers who try to register their own model on your machine. An open Ollama port is not a hobby if it can pull a model you pay to run.

After the foothold, the behavior GTIG tables is familiar cloud crime with a new last step. Reconnaissance against repos, buckets, containers, and access logs. Unauthorized IAM users and service accounts. Admin policies attached to roles that did not have them. GPU instances requested, quotas raised, pay-as-you-go model APIs turned on in projects where those APIs had been off. One actor used a burn-and-rotate pattern, creating and deleting short-lived service accounts to stay under rate limits. Privileged service accounts were enough to persist. They did not need a traditional malware implant to keep the bill running.

The LiteLLM case is the one that belongs to teams who think a gateway makes the key safe. Mandiant says a follow-on actor took virtual keys, the limited credentials a proxy issues, and abused CVE-2026-40217, a user-id rebind on the key regenerate endpoint, to turn those limited keys into full proxy-admin credentials. That is a sandbox escape in the gateway, not in the model. Patch LiteLLM. Rotate every virtual key and every admin credential that gateway could mint. Do not look for a walkthrough. The vulnerability class is "a limited key could become the master key." Your check is the version of the gateway and the date you last rotated.

What they do with the quota

Inside the account, the mission is to call models and to sell that ability. GTIG describes actors enabling managed model APIs, creating more keys, and subscribing victim accounts to pay-as-you-go plans on expensive models. One cluster, which they track as UNC6993, ran past $100,000 in charges after turning services on. Another path is to provision GPU machines and run the attacker's own tooling there, including agents that then look for more to steal in the same network. The quota is both the product and the funding for the next step.

The resale layer is what they call LLM mixers. A mixer is a store that does not own the models. It routes a buyer's traffic through a pool of stolen keys and stolen compute, spreads the volume so no single victim looks like the whole store, and charges the buyer a price that only works if the upstream bill is someone else's. GTIG says the labor is split. One group gets the initial access or buys it from infostealer logs. Another runs the store. Some recruit people to open free trials. Free-trial credentials were cheaper on the forums than enterprise ones. Prices they observed per million tokens ranged from about $0.10 to $25, depending on the model and the kind of key. One vendor, APIMart, raised both input and output prices fivefold once a prompt passed 100,000 tokens. Others, such as JetNight, sell monthly subscriptions instead of per-token rates.

The November 2 note is sharper about one store. An actor passed nearly one million dollars in token cost to a victim while powering a service at jet-night.com, after more than 500,000 calls in under eight hours. GTIG says Jet Night advertised an unlimited plan around $60 a month for frontier-style access that would normally cost more, and that the service posted outages on Telegram when keys were revoked. That is what revocation looks like from the store's side: an outage. From your side it looks like a bill you already owe. Revoke early. The store's downtime is not your refund.

They also describe lookalike relay domains and open-source gateways, including new-api and a proxy that turns a command-line session into an API, used to pool keys. One persona, D1337-KAIROS, fronted a Node proxy in a victim cloud account, with a container that called a model and then ran the model's commands locally, plus search and a sandbox. A key that can do that is not a chat key. It is a remote admin channel that happens to speak a completion protocol. If your agent framework key can also start compute, you have issued that channel.

A noisy second thief can be the only reason you notice. In one case two actors shared the same leaked credentials. The first enumerated model services and looked at cost data. The same day, a set of addresses with one user agent, likely a proxy or botnet, used those credentials to make millions of calls in about twenty-four hours. The victim revoked the key because of the noise, before the first actor could get their own queries running. You should not count on a second criminal to trip your alarm. The alarm should be the first unusual enablement of a model API, and the first spike, not the millionth call.

What to watch, in the order the bill arrives

Mandiant's detection advice is blunt. In several incidents the model API was not even enabled when the attacker arrived. The stolen key was allowed to turn it on. Treat "someone enabled Bedrock" or "someone enabled the Gemini API" in a project where that was off as a security event, the same day, not as a FinOps curiosity. Then watch for new keys, new service accounts, quota increases, GPU instance types you do not use, and calls from regions you do not serve.

Spend caps help and do not finish the job. A cap stops a weekend. It does not stop a key that was copied yesterday and is still under the cap at noon while it scouts. Combine the cap with an alert on the rate of spend and on the enablement event. A key per environment, unable to create other keys, is a smaller theft than a key that can administer the project. The virtual-key bug is the argument for that split: a limited key should not be able to become the admin key, and when a bug says it can, you patch before you trust the split again.

Keys in git are still the front door. A one-token probe will find them. Secret scanning on the repo and on the history, rotation when a scanner matches, and a ban on long-lived model keys in agent shells. An agent that needs a model gets a key that can call one model, from a broker, with a ceiling. The key does not live in the prompt and does not live in the repo. The crews GTIG quoted already tell their own operators not to paste keys into chat. Match them.

Session tokens are the MFA bypass. A phishing page that asks for a pasted JSON is not stealing a password. It is stealing a session that has already passed MFA. Train people not to paste tokens into anything but the real console, and shorten the token. A session that lasts a week is a product on a forum.

If you run a gateway, LiteLLM or anything like it, put it on the same patch clock as an internet-facing admin. The patch-exploitation note is why that clock is measured in days. A gateway bug that upgrades a customer key into an admin key is a perfect n-day: the patch tells the attacker exactly which check was wrong. Version the gateway. Do not expose its admin port.

The underground prices are a way to see your key the way a buyer does. A forum price of pennies per million tokens, or a sixty-dollar monthly plan that claims frontier models, only clears if the seller is not paying the real rate. You are the real rate. When a salesperson offers you model access far under the published price, through an unofficial gateway, assume the margin is someone else's invoice until they show a contract with the lab. Do not point a production workload at a mixer because the demo was cheap. The outages those stores post, when victims rotate keys, will be your outage too, and the prompts you sent will have transited a criminal proxy. Cheap inference with no vendor name is not a procurement win.

Inside your own cloud, the enablement event is the cheapest detector you are probably not using. Model APIs are often off in projects that do not need them. An attacker who wants the quota has to turn them on, subscribe a pay-as-you-go plan, or request GPU quota. Each of those is an API call your audit log already stores. Page a person on that call. Do not fold it into a weekly cost email. Mandiant's point about the spike arriving in hours is a point about paging, not about a better dashboard. A dashboard is where you confirm the page. It is not the page.

Keys issued to agents deserve the same alarm as keys issued to people, because the agent does not have a laptop you can remote-wipe. When an employee leaves, you disable the account. When an experiment ends, disable the key the same day. Orphan keys are what a mixer wants: valid, unused, and unwatched. A monthly job that lists keys older than the experiment they belonged to is the control. If you cannot list them, you cannot rotate them, and the jacking report's "revoke access" step is a step you will fail. Put the list in the same place as the invoice owner, so the person who sees the spike can also see the key. A finance channel that knows the dollar amount and a security channel that knows the key name are how the hour gets wasted on a bridge call. One owner, both facts.

What to do if the graph already spiked

  1. Revoke the key and the session before you write the postmortem. Disable model APIs that should not have been on. The store's resale stops when the key stops. Your liability for the hours already billed is a finance conversation. The next hour is a security action.
  2. List every key the compromised identity could create. Attackers create new ones. Rotating the key you know about and leaving the key they minted is how the bill resumes under a different name.
  3. Check for new service accounts, new admin policies, new GPU instances, and quota tickets. GTIG says persistence often looked like IAM, not malware.
  4. Patch the model gateway and rotate virtual keys if you run LiteLLM or a similar proxy. Confirm the version that fixes CVE-2026-40217. Do not keep a limited key that the unpatched regenerate path can promote.
  5. Look at git history and CI logs for the key. If it was public, assume it was probed with a one-token call long before the spike.
  6. Turn on an alert for API enablement and for a spend rate you would notice in an hour, not in a month. Test the alert.
  7. If customer data could have been in the prompts the attacker sent, or in logs the new admin key could read, that is a disclosure question for counsel. The jacking report is about quota. A key that can also read stores is a different incident sharing the same credential. Say which one you have.

LLM development that ships with one shared key and a monthly bill review is the product these mixers are built to resell. Split the key. Watch the enablement. The spike is the business model, and it is faster than your close of month.

If this is your app, send the repo.

A 30-minute scope call. We read what is live, name what is dangerous, and quote a fixed scope.

How Techtribe finds the key a mixer would resell

We list every model key and every gateway, including the ones an agent or a CI job holds, and we separate a key that can only call one model from a key that can turn APIs on. You get the rotation list, the spend alert that fires in an hour, and the gateway version. We do not wait for the invoice, and we do not describe how to abuse a virtual key.

A single leaked key can be a fixed pass. A product that mints keys for every new tool is a Dedicated Team if the review has to stay. The scope call is 30 minutes with a founder, and it is where we say which you are in.

Bring the last model invoice and where the keys live. If you have never set an alert on spend, say that. It changes the order of the call.

The fixed scope for taking a generated app to production is a page per tool: Lovable, Bolt, v0, and Replit.

Questions

What is LLM jacking?

It is the name Mandiant and Google's Threat Intelligence Group use for intrusions that steal access to a company's models or cloud GPUs and consume that quota, often to resell it. They contrast it with cryptojacking: the bill spikes in hours, not over weeks.

How expensive did Mandiant say it gets?

Across incidents in 2026, Mandiant reported fraudulent charges above $100,000 in a single day, and cumulative charges reaching $1 million at some organizations before access was cut. Those are their case figures, not a forecast for every stolen key.

What is an LLM mixer?

GTIG describes illicit storefronts that route buyers through pools of stolen credentials and compute so the buyer gets model access and the victim gets the invoice. Advertised prices they saw ran from about $0.10 to $25 per million tokens, with free-trial keys cheaper than enterprise keys.

What is CVE-2026-40217 in this context?

Mandiant describes it as a bug in LiteLLM's key regenerate path that let an attacker turn a limited virtual key into full proxy admin credentials. The response is to patch the gateway and rotate keys. This article does not describe how to perform that step.

Book a free call

Have a vibe-coded app in production?

30-minute free call. We'll read what you have, name what is dangerous, and quote a fixed scope.