All articles

NewsNovember 2, 20265 min read

LLMjacking: stolen AI access costs up to $100,000 a day

Google and Mandiant detail the LLMjacking supply chain: stolen API keys and cloud credentials, access resold via LLM Mixers, charges of up to $1 million.

Bar chart with a falling arrow cracking the floor, on an orange background with binary code

On 2 November Google Threat Intelligence Group (GTIG) and Mandiant published a report on LLMjacking, the theft of access to the large language models (LLMs) that companies use in the cloud. In the incidents Mandiant handled in 2026, fraudulent charges exceeded $100,000 in a single day and, at some organisations, reached $1 million before access was revoked. Victims were in financial services, retail, technology and transportation, across North America, South America and Asia Pacific.

The difference from cryptojacking is speed: mining consumes resources gradually, while LLMjacking can generate huge cost spikes within hours. According to BleepingComputer, once access is resold traffic can exceed 1,000 requests per second, which the provider records as legitimate consumption and bills to the victim.

A two-tier supply chain

GTIG describes a division of labour: on one side, actors who obtain initial access to cloud environments and AI services; on the other, those who monetise it through so-called LLM Mixers, gateways that pool stolen credentials, spread requests across accounts, hide where the access comes from and resell model usage to underground subscribers. Some cases from the report:

  • Jet Night offers an “unlimited” plan with frontier models for about $60 a month, against $100–250 for a comparable legitimate service: GTIG believes it can only be profitable with stolen keys. An actor feeding it ran more than 500,000 API calls in under eight hours with third-party keys, leaving the victim with nearly $1 million in costs. The outages the service announced on Telegram likely coincide with victims revoking keys.
  • UNC6993 broke into a cloud instance through a misconfigured service account key, created additional keys and enabled services in the victim’s projects, running up $100,000 in charges. According to CSO Online it pooled keys with open source gateways such as new-api and gcli2api. The group also ran a fake portal for “joining a group of LLM users” that tricked visitors into pasting the JSON containing an active access token, bypassing passwords and MFA.
  • According to Help Net Security, one attacker registered a lookalike domain posing as an internal host of the victim to funnel its AI quotas into the newapi.pro resale service.
  • Another actor, The Register reports, signed the victim up for pay-as-you-go agreements for premium models and created and deleted short-lived service accounts to get around rate limits (burn and rotate), while watching the billing page.

The attack chain

Initial access. Attackers skip chatbots and go straight for model resources: LLM endpoints exposed to the internet or valid credentials, such as API keys harvested from public GitHub repositories, IAM credentials, service account keys, GitHub personal access tokens and session tokens.

Validation. Stolen keys are tested with minimal requests, even capped at a single token, below the thresholds of consumption-based alerts.

Reconnaissance and persistence. Once inside, attackers hunt for more credentials in repositories, logs, databases and secret managers. In one case described by Help Net Security, the intruder queried data warehouse tables for passwords and sensitive columns and ran storage operations consistent with data theft. Backdoor accounts were observed, one disguised as an internal billing optimisation tool, along with injected SSH keys.

Services switched on by the intruder. In several incidents the AI services were disabled at the time of the compromise: the attackers enabled them with the stolen keys. According to BleepingComputer, one actor also used the victim’s infrastructure to host public reverse proxies exposing endpoints compatible with multiple models, resold under the “D1337-KAIROS” persona.

Vulnerable LLM gateways. The sources agree that LiteLLM was abused, but not on the details: Help Net Security says CVE-2026-40217 let attackers turn ordinary access keys into administrator credentials; BleepingComputer describes the same CVE as server-side code execution via the custom-code guardrail testing endpoint; CSO Online refers to a user ID rebinding flaw that elevates virtual keys to administrative ones, without a CVE number.

In MITRE ATT&CK terms, the activity maps mainly to T1496 (Resource Hijacking), T1078.004 (valid cloud accounts), T1552 (unsecured credentials), T1136.003 and T1098 (creating and manipulating cloud accounts) and T1190 for the gateway exploitation.

Impact for European and Italian companies

The sources do not mention European victims, but the pattern does not depend on geography: any company with cloud projects, service account keys or API keys in code is a potential target, even if it has never enabled AI services. For entities in scope of NIS2, an incident that causes or could cause severe financial loss is among the significant incidents that must be reported. Where the intruder explored databases and storage, as in the incidents described, a personal data breach under the GDPR must also be assessed.

What to do now

  1. Set spending and usage limits on model APIs and budget alerts on every cloud project.
  2. Deny by default access to cloud LLM services where it is not needed.
  3. Reduce the privileges of users and service accounts, and scope each API key to its intended models, workloads and endpoints.
  4. Centralise secrets on a platform with rotation, and automatically scan code and repositories for keys and secrets.
  5. Use phishing-resistant MFA and warn users never to paste tokens or JSON payloads into external pages.
  6. Enable and monitor audit and inference logs: AI service enablement, new keys or accounts, minimal validation requests, consumption spikes, attempts to disable logging.
  7. Update LiteLLM and other gateways, keep them off the internet and look for unauthorised AI services (shadow AI).

How ARXDOME helps

ARX Expose shows internet-facing services and recognised products, including an AI gateway or endpoint published by mistake, and flags company accounts involved in breaches. ARX Shield lists open vulnerabilities by machine and by package on monitored servers, including those hosting an LLM gateway, while ARX Sentry groups server events such as credential use, account changes and persistence.

Last updated:

Any questions?

Let’s talk about your company’s needs, with no obligation.

Contact us

Related articles