Criminals Are Now Reselling Your Stolen AI Credentials as a Monthly Subscription
Cybersecurity

Criminals Are Now Reselling Your Stolen AI Credentials as a Monthly Subscription

Google's own threat intelligence team found attackers draining enterprise LLM accounts and reselling the access through underground subscription services. One victim absorbed nearly a million dollars in token charges before anyone noticed.

PublishedNovember 20, 2026
Read time5 min read
Share

What LLMjacking actually is, and why it scaled into a business

LLMjacking describes the hijacking of enterprise large language model access, where attackers obtain valid API keys or cloud credentials and use them to run their own workloads on someone else's account, and someone else's bill. Google Threat Intelligence Group and Mandiant, in a report covered by multiple outlets on November 3, describe two distinct operational clusters coordinating to compromise and monetize enterprise LLM infrastructure at scale, which is a notably more organized structure than the opportunistic credential theft this category of crime started as.

The monetization model is the part that should concern enterprise security leaders most. Stolen initial access gets sold to what the industry calls LLM mixers, services that resell unauthorized model access to subscribers who want cheap, high-volume LLM usage without their own API relationship. One such service, Jet Night, reportedly sells unlimited access plans for sixty dollars a month, and Google found it had passed nearly one million dollars in token costs onto a single victim organization. Sixty dollars a month for the buyer, up to a million dollars for the victim footing the actual bill: that spread is the entire business model.

How the credentials actually get stolen

Attackers are not relying on one single technique. Google's report describes scanning for publicly exposed LLM endpoints, harvesting valid API keys left on unprotected machines, and exploiting compromised cloud credentials, service account keys, GitHub tokens, and active session tokens wherever they can be found. One tracked cluster, UNC6993, is specifically exploiting CVE-2026-40217, a flaw in LiteLLM that allows arbitrary code execution through a custom-code guardrail testing endpoint, a detail worth noting because it means the vulnerability sits inside AI infrastructure tooling itself, not just in generic cloud credential hygiene.

Once inside, the pattern repeats across incidents: attackers search for additional credentials to expand their foothold, then create backdoor accounts or inject SSH keys to maintain persistence well beyond the initial compromise. In some cases, victims' own cloud infrastructure was used to host the unauthorized proxy services and agentic workloads being resold to paying subscribers, meaning the victim organization was not just paying for the attacker's usage, it was hosting the attacker's criminal infrastructure on its own compute.

The numbers that should reach the CFO, not just the CISO

Google's report documents request volumes that suddenly exceeded one thousand requests per second on compromised accounts, and one actor made more than five hundred thousand LLM API calls in under eight hours. These are not subtle anomalies that require sophisticated anomaly detection to catch, they are order-of-magnitude spikes against any organization's normal usage baseline. The fact that some victims absorbed over one hundred thousand dollars in a single day before noticing suggests that AI spend monitoring, where it exists at all, is often disconnected from security alerting entirely, treated as a finance dashboard problem rather than a fraud detection one.

That gap between finance and security monitoring is itself the vulnerability here. A sudden thousand-dollar spike in a cloud compute bill triggers alerts in most mature organizations. A sudden hundred-thousand-dollar spike in an AI API bill apparently did not, in multiple cases, trigger anything until the bill arrived. Closing that gap requires treating AI token spend as a security-relevant signal with real-time alerting thresholds, not a monthly finance reconciliation line item reviewed weeks after the damage is already done.

Why this looks like a maturing criminal market, not a one-off scheme

The presence of named, branded relay services like Jet Night, with advertised pricing and marketed access tiers, is itself significant. A sixty-dollar-a-month unlimited access plan is a product, with a price point, a value proposition, and presumably some form of customer support for the criminals buying the access. That level of productization typically only emerges once a criminal technique has moved past the experimental phase and proven reliably profitable across enough victims to support dedicated infrastructure and ongoing operations rather than one-off opportunistic theft.

Google's description of two distinct operational clusters coordinating to compromise and monetize enterprise LLM infrastructure reinforces that reading. Separate clusters focusing on different stages, credential harvesting on one side, access resale and monetization on the other, suggests a division of labor similar to what security researchers have long documented in more mature cybercrime markets like ransomware-as-a-service. If LLMjacking is following that same trajectory, enterprise defenders should expect the volume and sophistication of these attacks to keep increasing rather than plateau.

Google's stated advice, deny all permissions that are not explicitly required, including access to cloud LLM services, is a textbook least-privilege principle that has been standard cloud security guidance for a decade. The fact that it needs restating specifically for AI services suggests many organizations rolled out LLM access with looser permission scoping than they would ever accept for a production database or payment system, treating AI infrastructure as experimental and therefore less deserving of the access controls applied everywhere else.

The fuller defense list, centralized secrets management with key rotation, phishing-resistant multi-factor authentication, monitoring for shadow AI usage, explicit API spending limits, and enabled cloud audit and inference logging, is not novel security advice in isolation. What is new is the urgency of applying it specifically to AI infrastructure, which many security teams still treat as an extension of existing cloud governance rather than as its own distinct attack surface with its own monetization economy already built around exploiting it.

Tagged#news#security#cybersecurity#breach#cisa#ransomware#zero-day#supply-chain#ai-security#llmjacking#google-gtig#mandiant#unc6993#litellm-cve#jet-night#ai-token-theft