There was no ransom note. No files were locked, no data was dumped on a leak site, no one called the executive team at two in the morning. The first indication that something had gone wrong, in several of these cases, was an invoice.
Google Threat Intelligence Group and Mandiant published research on November 2, 2026 on a pattern they call LLMJacking: attackers get hold of cloud credentials or AI API keys, then run their own workloads on the victim’s paid model access and leave the victim holding the bill. In separate incidents, the fraudulent usage topped $100,000 on a single day. At some organizations the running total hit $1 million before the access was shut off.
What makes this worth ten minutes of your attention is not the dollar figure. It is that nothing novel was required to produce it. This is a credential hygiene problem in an AI costume.
What Happened
Mandiant investigated multiple intrusions during 2026 spanning several cloud and model providers. Victims were spread across North America, South America and Asia-Pacific, in financial services, retail, technology and transportation.
The entry point was rarely exotic. The researchers describe the dominant route as cloud identity material that was either exposed or stolen: IAM credentials, service account keys, GitHub personal access tokens and live session tokens. One actor simply scraped corporate model API keys out of public GitHub repositories, then tested which ones still worked by sending requests capped at a single token, quiet enough that volume-based alerting in the victim’s EDR and SIEM never fired.
One tracked group, UNC6993, got into a cloud instance in mid-2026 through a misconfigured service account key and used that position to reach an LLM API. The same actor also ran a page dressed up as an application portal for joining a group of model users. Visitors were walked through pasting a raw JSON payload containing a live access token, which the attacker then replayed against the backend. Passwords and MFA were not defeated. They were stepped around, because a valid session token does not need either.
From there the playbook is ordinary cloud intrusion work: enumerate services, clone repositories looking for hardcoded secrets, query databases for more credentials, attach administrative policies, inject SSH keys. One actor created backdoor service accounts named to look like internal billing optimization tools, which is a detail worth admiring briefly before moving on. To stay under rate limits, another used what Google describes as a “burn and rotate” approach, spinning up and deleting short-lived service accounts while watching the cloud provider’s own billing console to track the damage.
The money comes from resale. Google documents an emerging layer of operators it calls LLM Mixers, which pool stolen keys and resell access through underground storefronts. One such service, Jet Night, was powered in part by an actor who pushed over 500,000 model API calls in under eight hours and passed close to a million dollars in token costs to the victim. Jet Night advertised an unlimited plan for roughly $60 a month, against a market rate the researchers put at $100 to $250. The economics only work if someone else is paying, and someone else was.
One finding deserves to be read twice: in several of the incidents, the LLM services were not even turned on when the attacker arrived. The stolen credentials had enough permission to enable them.
Why This Matters for Businesses
Cryptojacking taught a generation of IT teams to watch for a slow, unexplained rise in compute spend. This is not that. Google’s point of comparison is explicit: where cryptomining bleeds a budget gradually, hijacked model access can spike a cloud bill within hours.
That compresses the detection window to something most mid-market organizations are not built for. Monthly invoice review is a fine financial control and a useless security control when the loss event takes an afternoon.
There is a second, quieter cost. Every one of these cases involved a credential with real permissions in a production cloud environment. The attacker who can enable a model API can usually also read your secret manager, clone your repositories and create accounts. The bill is the loud part. The access is the expensive part.
And most organizations have more of this material than they think. Over the last two years, AI keys have been issued by developers testing an idea, by a marketing team wiring up a content tool, by a vendor integration that someone approved and no one owns. They were created as experiments. They are now production credentials, and almost none of them went through the identity governance process that a new employee account would trigger automatically.
What Organizations Should Consider
The encouraging part of this story is how conventional the fixes are. Google’s own remediation guidance leads with identity and access management and supply chain hygiene, not with anything AI-specific.
A few questions worth putting to your IT team or provider this month:
Who currently holds an AI or model API key, and where is it stored? If the answer takes more than a week to produce, that is the finding. Google recommends centralized secrets management with automated rotation, precisely so a leaked key has a short useful life.
Are service account keys and personal access tokens treated as privileged credentials? They frequently are not, because no human logs in with them. The research shows attackers treating them as the fastest path into a cloud estate.
Do we scan source code for secrets automatically? Harvesting keys from public repositories was one of the observed access routes. Automated secret scanning across GitHub, GitLab and equivalents is a solved problem with free and paid options.
Are there spending caps and per-model quotas on our AI services, and who gets the alert? Budget alerts and granular quotas appear in Google’s mitigation list for a reason. A spend alert that reaches a security inbox within the hour turns the invoice from a post-mortem into a detection.
Are our API keys scoped? The guidance suggests segmenting keys by environment and by model endpoint, so a token for a low-cost internal utility cannot be turned around and used to invoke premium frontier models.
Would we notice a model service being switched on? Given that attackers enabled services that had been dormant, unauthorized enablement of an AI API is a clean, low-noise signal that something is wrong.
Is our model gateway in the patch program? Google cites exploitation of CVE-2026-40217 in LiteLLM, where an actor elevated ordinary virtual keys into full administrative credentials. Middleware that routes AI traffic is infrastructure and should be patched like it.
None of that requires a new product category. It requires deciding that AI credentials are credentials.
How ISM Can Help
An AI Readiness & Automation engagement is the natural place to start, because the first deliverable is an honest inventory: which model services are in use, who issued the keys, where they live and what spending limits exist. Most organizations find at least one key they had forgotten about.
From there, Identity & Access Management work brings service account keys, personal access tokens and session lifetimes into the same governance you already apply to user accounts, including phishing-resistant MFA on the consoles and developer platforms that issue these credentials in the first place.
Cloud Solutions & Microsoft 365 reviews cover the configuration side: least-privilege roles, exposed endpoints, logging that is actually enabled, and quotas and budget alerts set at a level where someone hears about an anomaly the same day.
And Cybersecurity & Risk Management is where the cloud bill stops being purely a finance artifact and becomes a monitored signal, alongside detections for unauthorized service enablement and unusual key creation. Clients on Insight Managed Defense and Insight Cyber Command get this stitched into ongoing monitoring rather than handled as a one-time project.
Talk With Our Team
If you have adopted AI tooling faster than you have governed it, you are in extremely normal company, and the cleanup is smaller than the headline number suggests. We are happy to walk through where your model keys live, who can mint new ones, and what a runaway bill would look like before it becomes one. Get in touch and we will start with your environment rather than a slide deck.
If you would rather get your own numbers together first, the QBR calculator is a reasonable way to frame what you are spending on cloud and IT today, which is usually the conversation that surfaces the forgotten subscriptions anyway.
Sources
- The Rise of LLMJacking: How Threat Actors Monetize Compromised Cloud and Model Resources, Google Cloud (Google Threat Intelligence Group and Mandiant)
- Google Warns LLMjacking Attacks Are Costing Some Enterprises Up to $1 Million in Cloud Charges, The Hacker News