Mandiant: LLMjacking Runs Up $100,000 a Day on Stolen Cloud Keys, and Logs Are Off by Default
Security / news
Mandiant: LLMjacking Runs Up $100,000 a Day on Stolen Cloud Keys, and Logs Are Off by Default
Google's Nov. 2 report says single-token test calls evade volume alerts, and it describes the LiteLLM bug differently from BleepingComputer.

Google's Threat Intelligence Group and Mandiant say attackers who steal cloud credentials are running up more than $100,000 a day in fraudulent AI charges on a single victim, with cumulative bills at some organisations reaching $1 million before access was revoked. The finding comes from The Rise of LLMJacking, published on November 2, 2026 and written by nine named Google authors including Christopher Romano and Mike Lombardi.
The post covers incidents Mandiant handled in 2026 at victims in North America, South America and Asia Pacific, across financial services, retail, technology and transportation. It does not give a total number of incidents, so the dollar figures describe the worst cases, not an average.
How the credentials are stolen
Exposed or stolen cloud identity tokens were the dominant way in: IAM credentials, service account keys, GitHub personal access tokens and session tokens. Google also describes attackers probing publicly exposed model servers such as Ollama, calling its /api/create route to register models, and harvesting corporate LLM API keys from public GitHub repositories.
One cluster, tracked as UNC6993, used a session-hijacking page that asked users to paste a raw JSON payload containing a live access token, which bypasses both the password and multi-factor authentication. Another actor used cloud federation tokens to mint fresh access keys and sidestep MFA.
The detection gap: one token at a time
The attackers confirmed that stolen keys worked with requests capped at max_tokens=1. Google says those single-token calls evaded volume alerts in EDR and SIEM tools. In several incidents the victim had never enabled LLM services at all; the stolen key switched them on, and the bill grew "in just minutes or hours".
Logging is a second gap. On Google Cloud, inference operations are recorded in Data Access logs, which are disabled by default because of volume and cost. A victim who never turned them on has no record of which prompts were sent. Google's guidance is to alert on ServiceUsage.EnableService for AI Platform services, on calls to PredictionService.GenerateContent, and on any attempt to delete the model invocation logging configuration on AWS Bedrock.
Who buys the stolen access
The post describes "LLM Mixers", gateways that "proxy, aggregate, and resell unauthorized LLM access to underground subscribers". UNC6993 used the open source tools new-api and gcli2api to pool stolen keys and balance load. A service called Jet Night sold an unlimited plan for about $60 a month, against a typical $100 to $250, and Google believes it runs on stolen keys. Its advertised outages are likely key revocations.
BleepingComputer's summary adds that one actor made more than 500,000 API calls in under eight hours, and that Jet Night passed nearly $1 million in token costs to one victim.
- Jet Night unlimited plan (stolen keys, per Google)60 USD
- Typical plan, low end100 USD
- Typical plan, high end250 USD
Source: Google Threat Intelligence Group, The Rise of LLMJacking, November 2, 2026
The LiteLLM bug, described two ways
Google reports that a follow-on cluster exploited CVE-2026-40217 in the LiteLLM gateway's /key/regenerate endpoint to turn standard virtual keys into PROXY_ADMIN credentials. BleepingComputer describes the same CVE as arbitrary code execution through a custom-code guardrail testing endpoint. The two accounts differ and The Terminal has not seen the CVE record, so the vulnerability class is unresolved. Neither source gives a severity score or a patched version in the text reviewed.
What defenders can do
Google's recommendations are specific: central secrets management with automated rotation, an explicit deny on LLM services for principals that do not need them, API keys scoped by model and environment, phishing-resistant MFA, and budget alerts with per-model quotas. Its own phrase is to "deny all permissions that are not explicitly required, including access to Cloud LLM services."

A spending cap is the cheapest control on the list, and it is the one that would have limited a $100,000 day. The same logic applies to self-hosted models: a server reachable from the internet, like Tencent's LoopForge exposing a Docker socket, is one request away from becoming someone else's endpoint, and open-weight releases such as Inkling 2 Small make self-hosting more common. Fireworks AI has linked its own incident to a group selling tokens from unauthorised model access, which it also calls LLMjacking; its November 9 bulletin is still marked "Active monitoring".
Sources
More in Security
- 01Java4Shell Scores 9.2 From Its Finder and 8.1 From OpenJDK, and Needs an AVX-512 CPUCalif says a 65-character Base64 string overwrites 62 bytes of Java heap and reaches code execution in Keycloak, Jenkins and Jira. A second security firm says the source shows only a crash.
- 02CISA Gave 9 of 16 November KEV Entries a Three-Day Deadline With Forensic Triage, and 7 Got 14 DaysThe catalog's own due-date and forensicTriage fields show two tiers under BOD 26-04. Sorting November's additions by tier shows which exploited bugs CISA treats as already-compromised-until-proven-otherwise.
- 03Project Zero Cuts Its Disclosure Deadline to 60 Days on Feb. 2, 2027, Then 30 in 2028Google's bug-hunting team is dropping the 30-day patch soak period and counting every deadline from the day it reports, citing AI-assisted exploitation.
- 04Google Finds exploit_agent Diffing Patches and Firing Exploits With No Operator in the LoopGTIG's November 16 tracker reports the first fully automated discovery-to-exploit pipeline it has seen, found on an exposed staging server, plus 22 zero-days in August.