Summary
Google Threat Intelligence and Mandiant are warning about LLMjacking: attackers stealing a company's access to AI models and leaving the company to pay the bill. Some victims saw more than $100,000 in fraudulent charges in a single day.
Key points
- Google Threat Intelligence Group and Mandiant observed fraudulent AI charges above $100,000 in a single day, with some organizations reaching $1 million before access was revoked.
- The most common way in was stolen or exposed cloud identity credentials: service account keys, GitHub personal access tokens, session tokens and API keys left in public code.
- Stolen access is resold through "LLM mixers," including one service advertising unlimited frontier-model access for about $60 a month.
- Spending caps, keys scoped to specific models and environments, centralized secret management and billing alerts limit how far a stolen key can go.
Attackers no longer need to steal your data to profit from a breach. Your AI bill will do.
Google Threat Intelligence Group (GTIG) and Mandiant have published research on a growing form of resource theft they call LLMjacking: stealing an organization’s access to large language models and letting that organization pay for the usage. In individual incidents, Google observed fraudulent charges above $100,000 in a single day, and some organizations reached $1 million in cumulative charges before the access was shut off. (Google Cloud)
If that sounds like cryptojacking, it should. The difference is speed. Cryptojacking tended to creep up over weeks. AI usage is billed per token, premium models are expensive, and an attacker with a high quota can run up a five or six figure bill in hours.
The Way In Is Ordinary
Nothing about the initial access in Google’s research is exotic. The most common entry point was exposed or stolen cloud identity credentials: IAM credentials, service account keys, GitHub personal access tokens and active session tokens. Other actors went straight for LLM API keys left in public code repositories. (The Hacker News)
One group Google tracks as UNC6993 compromised a cloud instance through a misconfigured service account key, used it to reach an LLM API, then created additional API keys and enabled more services inside the victim’s cloud projects. Charges passed $100,000.
Another actor scraped LLM API keys from public GitHub repositories and tested each one with a request capped at a single output token. Requests that small are easy to miss in volume-based monitoring. (SiliconANGLE)
Stolen Access Is Now a Product
The newer development is the business model. Google describes “LLM mixers,” gateways that pool stolen keys, route customer prompts through hijacked accounts and sell the access through underground storefronts. One service, Jet Night, advertised an unlimited plan with alleged frontier-model access for about $60 a month. In one case, a mixer using stolen third-party keys made more than 500,000 LLM API calls in under eight hours.
That division of labor matters. The person who steals the key and the person who spends it are often different people, which means a leaked key can be monetized within hours of exposure.
Treat an AI API Key Like a Corporate Credit Card
Most organizations already have rules for corporate cards: limits, named owners, statements someone reviews and a fast way to cancel. AI API keys deserve the same treatment, and in many environments they get almost none of it. Keys are created by developers during experiments, pasted into config files and scripts, and left active long after the project ends.
Google’s guidance and the broader reporting point to a short list of controls:
- Set spending caps and billing alerts on every AI provider account and cloud project with model access. This is the cheapest control and the one that limits damage when everything else fails.
- Scope keys narrowly to specific models, environments and usage limits, so a key issued for a low-cost model cannot be used to call expensive frontier models.
- Centralize secrets in a managed vault and rotate them on a schedule, rather than storing keys in code, notebooks or environment files.
- Run secret scanning on code repositories, including personal and public repositories your developers use for work.
- Review service account permissions and remove the ability to create new keys or enable new services where it is not needed.
- Watch the logs for unusual model usage, new keys, newly enabled AI services and very small validation requests.
- Patch AI gateways and proxies such as LiteLLM promptly. Google’s research cites exploitation of a LiteLLM vulnerability as one escalation path.
The Governance Question
LLMjacking is a technical problem with a governance root. Many organizations cannot answer basic questions about their AI footprint: which teams have model access, who owns each key, what the spending limits are and who would notice an anomaly.
Who in your company would notice a $100,000 AI charge before the invoice shows up? If the honest answer is “finance, at month end,” the controls above belong on this quarter’s roadmap.
Get Ahead of AI Credential Risk
Cyber Defense Advisors helps organizations inventory their AI usage, test AI-enabled systems and put governance around the credentials that power them. Learn more about our AI security testing and AI security governance services, or schedule a conversation with our team.