On this page (7 sections)

Google has warned that attackers are hijacking enterprise access to large language models (LLMs) and running up cloud bills on the victim's account. According to reporting by The Hacker News and Infosecurity Magazine, Google Threat Intelligence Group and Mandiant observed fraudulent charges above $100,000 in a single day, and cumulative charges reaching $1 million at some organizations before access was revoked.

If your website, store or application calls an LLM through an API key, or your team keeps cloud credentials in code repositories, this applies to you. Note that no original Google report or vendor advisory was available to us; the details below come from press coverage of a report shared with The Hacker News, and we attribute claims accordingly.

What Google reported

LLMjacking is resource hijacking adapted to AI services. An attacker obtains an API key or cloud identity token, uses the victim's LLM quota and compute, and leaves the legitimate owner to pay. Google said that unlike gradual cryptojacking, it can produce large spikes in fraudulent charges within hours.

Per the reports, the activity ranges from low-level abuse, such as manipulating a customer support chatbot into doing unauthorized work like code generation, to direct targeting of backend model resources through exposed LLM endpoints or valid API keys. Infosecurity Magazine adds that Google sees a division of labor between actors who gain initial access and those who monetize it, including so-called "LLM Mixers", routing gateways that proxy, aggregate and resell LLM access to criminal customers. Google expects criminals to prioritize LLMjacking over cryptojacking going forward, according to that report.

How attackers get in

The Hacker News reports that the most common initial access vector was exposed or stolen cloud identity tokens. The reported routes include:

  • Phishing, source code leaks, misconfigured environments and exposed client-side code that reveal an API key.
  • IAM credentials, Service Account keys, GitHub Personal Access Tokens and active session tokens.
  • LLM API keys found in public GitHub repositories.
  • A hacked Service Account key with broad project permissions.
  • Compromised corporate IAM credentials used to generate fresh access keys, which sidesteps MFA safeguards.
  • A suspected web application vulnerability in a package running in a containerized cloud environment, following a stolen Personal Access Token used to reach private repositories.

One cluster, tracked as UNC6993, reportedly breached a cloud instance through a misconfigured Service Account key. Google said it also hosted a fake application portal that tricked users into pasting a raw JSON payload containing an active access token, which was then used to bypass password and MFA controls. The sources describe this as one actor's tactic; they do not say how widespread it is.

What happens after the foothold

Reporting says actors move quickly to escalate privileges, map the environment and find more credentials. Observed behavior includes cloning codebases to search for hard-coded secrets, pulling credentials from cloud secret managers, creating unauthorized IAM and service accounts, injecting SSH keys into compute instances, and creating backdoor service accounts disguised as internal billing optimization utilities. The Hacker News also lists exploitation of vulnerabilities in LLM routing software, naming CVE-2026-40217 in LiteLLM. The sources give no score, affected versions or fix for it, so check the LiteLLM project's own advisories if you run it.

Who is realistically exposed

  • Sites and apps that embed an LLM API key in front-end JavaScript or mobile code.
  • Teams that commit keys or service account JSON files to repositories, public or private.
  • Organizations with long-lived, broadly scoped service account keys or personal access tokens.
  • Anyone running LLM routing or gateway software exposed to the internet without tight access control.
  • Businesses with no spending alerts on their AI or cloud accounts, where a spike could run for hours unnoticed.

How to check your exposure

Start with places secrets tend to leak. This read-only search, run from the root of a local clone of your repository, lists files containing common key-like strings so you can review them by hand. Expect some false positives, and treat any real hit as compromised.

Search a local repository checkout for likely hard-coded secrets (run in the repo root)Shell
grep -rIn --exclude-dir=.git -E "(api[_-]?key|secret|token|private_key)" .

Then check your history, since a key removed from the latest commit may still sit in older ones. This lists commits that added or removed lines matching the pattern, so you can inspect them.

List commits touching key-like strings in git history (read-only)Shell
git log --oneline -S"api_key" --all

Finally, review your cloud and LLM provider billing and audit logs for usage you do not recognize, new service accounts, new access keys and unfamiliar SSH keys on compute instances.

What to do now

  1. Inventory every LLM API key, service account key and personal access token, and note who owns each.
  2. Revoke and rotate any key that was ever in a repository, front-end code, a ticket or a chat, then confirm the old one stops working.
  3. Move LLM calls to your server side so keys never reach the browser or app; proxy requests through your backend.
  4. Replace long-lived keys with short-lived credentials where your provider supports it, and scope each identity to the minimum permissions.
  5. Set budget alerts and hard quotas on LLM and cloud accounts so a spike triggers a page within hours, not at month end.
  6. Audit for unexpected IAM users, service accounts, access keys and SSH keys, especially anything named like a billing or cost tool.
  7. Enforce MFA on cloud and GitHub accounts, and train staff not to paste tokens or JSON payloads into unfamiliar pages.
  8. Keep LLM routing and gateway software updated, restrict who can reach it, and review the vendor advisory for CVE-2026-40217 if you use LiteLLM.
  9. Turn on secret scanning for your code hosting and run it on existing repositories.
Caution. Rotating a key breaks every application that still uses it. List dependents first, deploy the new key, then revoke the old one. If you suspect active abuse, revoke first and accept brief downtime.

Detection and monitoring

Because charges can climb by $100,000 or more in a day according to Google's reported figures, speed matters more than precision. Alert on request volume per key, usage from new regions or IP ranges, use of models you do not normally call, and sudden GPU or compute provisioning. Log administrative actions such as creating service accounts or keys, and review them regularly.

You can also check your public footprint. A security headers check and the free malware scanner will not find leaked cloud keys, but they help you spot other weaknesses in the same web properties. Outside-in checks cannot see inside your repositories or cloud accounts, so pair them with the internal reviews above; see what outside-in detection can and cannot see. For broader cloud-credential context, our coverage of residential proxy traffic on a government network shows how proxied access can obscure where traffic originates.

What is not yet known

  • The sources do not name the affected organizations or give a total victim count.
  • No primary Google report was available to us, so figures and tactics rest on press coverage.
  • There is no severity score or patch information for CVE-2026-40217 in the research.
  • It is unclear how much of the activity involves small sites versus large enterprises; the reports focus on enterprises.

Sources

This article is based on the public sources below. Details can change as vendors and researchers publish more; check the original advisories before acting on version numbers.

  1. Google Warns LLMjacking Attacks Are Costing Some Enterprises Up to $1 Million in Cloud Chargesthehackernews.com ·
  2. LLMJacking is Costing Some Businesses Up to $1minfosecurity-magazine.com ·
  • LLMjacking
  • API keys
  • cloud security
  • Mandiant
  • resource hijacking