Google warns of LLMjacking spike hitting firms hard
Mon, 16th Nov 2026 (Today)
Google Threat Intelligence Group has reported a rise in attacks that hijack corporate access to large language models, with cloud charges exceeding USD $100,000 in a single day.
Mandiant investigated multiple cases in 2026 across North America, South America and Asia Pacific involving companies in financial services, retail, technology and transportation. In some cases, organisations racked up fraudulent costs of as much as USD $1 million before access was cut off.
The attacks, described as LLMJacking, differ from older forms of resource theft such as cryptojacking because they can trigger sharp billing spikes within hours. The activity has affected a range of cloud and model providers and appears organised enough to show a split between groups that gain access and others that monetise it.
How access was taken
Attackers have focused on exposed model endpoints and stolen credentials, including identity and access management credentials, service account keys, GitHub personal access tokens and active session tokens. In several cases, they also harvested API keys left in public code repositories and used minimal test requests to check whether those keys still worked while trying to avoid detection.
One case involved a misconfigured service account key that gave a threat actor access to a cloud instance, followed by access to an LLM API. Another involved a page posing as an application portal that persuaded users to paste a raw JSON payload containing an active access token, allowing the attacker to bypass password and multi-factor authentication checks.
Elsewhere, attackers combined a stolen personal access token with the exploitation of a suspected web application flaw in a package running in a containerised cloud environment. In another incident, compromised corporate IAM credentials were used with cloud federation tokens to generate fresh access keys.
After entering a victim environment, attackers moved laterally by searching code repositories, cloud services, databases and storage systems for more credentials. Some queried metadata tables in cloud data warehouses for administrator passwords and other sensitive fields, while others cloned internal codebases and reviewed access logs for credentials tied to secondary cloud environments.
Persistence and privilege
The investigation also found repeated use of cloud permissions to escalate access and maintain a foothold. Threat actors created unauthorised IAM and service accounts, attached administrative policies to service roles and injected SSH keys into compute instances.
In one intrusion, a follow-on cluster used a vulnerability in LiteLLM gateway software, identified as CVE-2026-40217, to turn lower-level virtual keys into administrative credentials. That gave the attackers deep and persistent control over the compromised environment.
Once access had been established, the attackers sought to monetise it by generating new API keys, turning on model services and running large volumes of inference requests. One threat actor created and deleted short-lived service accounts in a "burn and rotate" approach while checking the cloud provider's billing interface to track the financial impact.
Resale market
Google Threat Intelligence Group said it is seeing signs of a broader underground market in stolen model access. It described "LLM Mixers" as illicit gateways that proxy, aggregate and resell unauthorised access to models.
In one example, a service called Jet Night is said to have used stolen third-party API keys to route tokens, making more than 500,000 API calls in less than eight hours. Investigators said the activity pushed close to USD $1 million in token costs on to the victim organisation.
The report linked other activity to open-source gateway and proxy tools, including new-api and gcli2api, which can be used to pool stolen keys and spread requests across compromised accounts. Researchers also observed custom relay infrastructure on deceptive lookalike domains designed to resemble internal company hosts.
Some attackers also tried to increase their computing resources by provisioning GPU-equipped instances in several regions and seeking quota increases through provider APIs. In one case, a threat actor used custom Node.js proxy bundles and deployed a containerised backdoor linked to the Manus agent framework in a compromised cloud environment.
Detection and response
Defenders should treat any unauthorised enablement of LLM services as a critical security event. In several incidents, the targeted services had not been enabled before the compromise, meaning the attackers first switched them on and then rapidly generated large bills.
The guidance calls for tighter credential management, least-privilege access, secret scanning in code repositories, routine configuration audits and stronger controls on API keys. It also recommends rate limiting, network allow listing, budget alerts, phishing-resistant multi-factor authentication and patching of model routing gateways and machine learning software dependencies.
For cloud monitoring, organisations are advised to watch audit logs for direct model execution, reconnaissance of available models and attempts to alter logging settings. The report also highlighted low-volume validation requests, such as single-token prompts, as a potential sign that attackers are quietly testing stolen credentials before launching large-scale abuse.
"As the LLMJacking ecosystem matures, we are also tracking a division of labor between threat actors establishing initial access to compromised environments and operators managing downstream monetization infrastructure, known as 'LLM Mixers'," said Christopher Romano, Codi Starks, Garrett Mahoney, John Wolfram, Josh Murchie, K Singh, Mathew Potaczek, Mike Lombardi and Rik Dolfing of Google Threat Intelligence Group.