← All articles

LLMjacking raids push cloud bills to $1 million

LLMjacking raids push cloud bills to $1 million

LLMjacking is turning stolen enterprise access to large language models (LLMs), AI systems that generate text, code and other content, into cloud bills that can reach $1 million. Google Threat Intelligence Group and Mandiant said attackers have caused fraudulent resource charges above $100,000 in a single day, while cumulative costs at some organisations reached $1 million before access was revoked.

The technique is a form of resource hijacking in which someone else consumes computing capacity while the account owner pays. Attackers obtain an application programming interface (API) key, a secret credential that lets software use an online service, through phishing, leaked source code, incorrectly configured environments or credentials exposed in software running on users' devices. They then use the victim's LLM account and its access to expensive AI processing. Affected organisations have been identified across North America, South America and Asia Pacific in financial services, retail, technology and transportation.

Google and Mandiant distinguish LLMjacking from lower-level abuse, such as manipulating a customer-support chatbot into generating code. LLMjacking reaches the underlying model resources directly, often through publicly exposed LLM endpoints, which are network addresses through which software requests model output, or valid API keys. Unlike cryptojacking, where attackers secretly use another system to mine cryptocurrency and may consume resources gradually, LLMjacking can produce a sharp rise in charges within hours.

The most common entry route observed was an exposed or stolen cloud identity credential. Examples included Identity and Access Management (IAM) credentials, which control what cloud users and systems may access, Service Account keys used by software rather than a person, GitHub Personal Access Tokens (PATs), and active session tokens that prove a user is already signed in. The threat actor tracked as UNC6993 entered one cloud instance through a misconfigured Service Account key and used it to reach an LLM API. It also operated a session-hijacking page disguised as a portal for joining a group of LLM users. Victims were directed to the legitimate LLM website and persuaded to paste a raw JSON payload, a structured block of data containing an active access token, into the fake portal. The token was then submitted to a backend API, bypassing password and multi-factor authentication (MFA), which normally requires an additional sign-in check.

In another incident, an unidentified attacker used a stolen PAT to enter private code repositories, then exploited a suspected web application flaw in a software package running inside a containerized cloud environment, where applications operate in isolated software units. Other routes included compromising employees' IAM credentials and programmatically creating fresh access keys to get around MFA, abusing a cloud Service Account key with broad project permissions, and collecting LLM API keys published in public GitHub repositories.

After entry, attackers escalated their permissions, created persistent access and searched internal systems for more credentials. Google observed them listing cloud services, code repositories and data stores; mapping services and identity policies across cloud regions; cloning sensitive code to find embedded secrets, configuration values and developer credentials; checking access logs for secondary cloud accounts; and querying databases for additional passwords. They also abused cloud permissions and virtual-machine settings, generated temporary credentials for highly privileged service accounts, and extracted infrastructure credentials, database passwords and third-party API keys from cloud secret managers.

Persistence methods included creating unauthorised IAM and service accounts, adding Secure Shell (SSH) access credentials to computing instances, and disguising backdoor service accounts as internal billing-optimisation tools. Attackers also exploited vulnerabilities in LLM routing software, including CVE-2026-40217 in LiteLLM, to raise their privileges. Their eventual goal was to monetise access to foundation models, the underlying AI models on which services are built, through token proxying, automated model streaming and graphics processing unit (GPU) hardware provisioning. Compromised environments were also used to host proxy endpoints and agentic workloads, in which AI software performs sequences of tasks with limited human input.

UNC6993 created additional API keys inside victim cloud projects to expand its activity, producing more than $100,000 in charges. Some attackers used a "burn and rotate" method to evade rate limits, repeatedly creating and deleting short-lived service accounts. Google said a more specialised criminal market has developed around LLM Mixers, illicit gateways that combine, route and resell unauthorised model access. One service, Jet Night at "jet-night[.]com", advertised unlimited tokens and alleged access to a frontier model for about $60 per month, allowing buyers to consume enterprise quotas while victims received the bills.

Symantec and Carbon Black, both owned by Broadcom, separately described a grey-market relay service active since at least early 2026. It primarily affected about a dozen organisations in Mainland China, plus multinational subsidiaries in South Korea, Taiwan and Malaysia. Sectors included industrial automation, electronics manufacturing, semiconductors, pharmaceuticals, shipping, tyre manufacturing, petrochemicals and data centre services. The service mainly appeared in markets where Anthropic and OpenAI do not operate, with installation initiated by employees seeking Western AI coding assistants through an unauthorised reseller.

Rather than supplying a modified or trojanised program, the service installed the genuine OpenAI Codex and Anthropic Claude Code command-line interfaces (CLIs), tools operated by typed commands. Soon after installation through npm, users were likely told to run a single command that downloaded a fileless PowerShell or bash setup script from the lookalike domain "vibe.aiok[.]me". The script changed "~/.codex/config.toml" for Codex or "~/.claude/settings.json" for Claude Code, disabled the execution sandbox, set Codex to "danger-full-access", allowed network access from the write sandbox, replaced users' genuine API credentials with relay-issued ones, and disabled telemetry and error reporting. It also checked the Windows locale to confirm that the computer was in the region served before completing setup.

Symantec suspects the operators were circumventing regional restrictions and reselling capacity below market rates, in a model resembling Jet Night. There was no evidence that the relay harvested employees' prompts or fabricated or changed model responses, but routing business requests through unapproved infrastructure and disabling safeguards creates clear security exposure.

Also available in: DeutschItalianoShqip