Mandiant reports multiple 2026 intrusions in which attackers resold access to organizations’ cloud AI resources and left the victims paying for it. Its November 2 investigation describes fraudulent charges exceeding $100,000 in a day, with cumulative charges at some organizations reaching $1 million. The technique is called LLMjacking: using someone else’s language-model access without authorization.
Evidence status: real intrusions reported by the investigating security vendor. The figures describe individual investigated incidents, not an industry-wide loss estimate. Bill Toulas’s BleepingComputer report summarizes the same research rather than independently confirming the bills.
An unused AI service can still become a billing risk
Google Threat Intelligence Group and Mandiant identify exposed or stolen cloud identities as the dominant initial access route in their investigations. These included service-account keys, developer tokens and active sessions. Attackers also harvested model API keys from public code repositories and looked for unprotected model endpoints.
In several incidents, the model service was not enabled when the attacker first entered. The stolen identity had enough permission to activate it. Restricting who may invoke an existing model therefore addresses only part of the risk. Permissions to enable services, create keys, impersonate other accounts and raise quotas also matter.
Some attackers expanded access into repositories, databases and secret managers. The report describes unauthorized accounts and added SSH keys used to retain access. Revoking one model API key may leave the identity or backdoor that created it intact.
The researchers call the resale gateways LLM mixers: services that pool stolen access and relay paying subscribers’ requests. In one investigation, they observed more than 500,000 model API calls in under eight hours associated with a service they identify as Jet Night. They attribute nearly a million dollars in token costs to that victim. This is a source-reported case, not evidence that every inexpensive AI relay uses stolen keys.
Limit authority before relying on usage alarms
The report supports these defensive priorities:
- Separate workload identities and model access. Grant only the services, models and environments each workload needs. Remove unnecessary rights to create credentials, enable services or assume more privileged identities.
- Rotate exposed credentials and investigate persistence. Review new users, service accounts, keys, SSH authorization and policy changes. Preserve relevant logs while containing unauthorized access.
- Monitor service activation and small credential checks. The investigators saw low-volume requests used to test keys before large-scale consumption. A detector that waits only for a spending spike can miss that preparation.
- Configure model quotas and appropriate spending controls. Check their actual scope, enforcement behavior and the permissions needed to raise or remove them.
- Verify inference logging separately from administrative logging. Records of account changes do not necessarily include model invocations. The report explains different logging requirements for AWS, Azure and Google Cloud.
Google’s billing documentation distinguishes alerts-only budgets from spend-cap budgets. Alerts-only budgets send notifications; they do not themselves stop requests. Its spend-cap documentation describes enforcement for eligible services, scoped to one project and one service. In-flight requests still complete, persistent resources can keep accruing charges, and caps reset with the next budget period. Verify coverage rather than treating a budget number as a universal ceiling.
The LiteLLM detail needs clarification
The investigation also reports abuse of a LiteLLM model gateway and names CVE-2026-40217. Its mechanism description conflicts with the published CVE record: the report describes key-regeneration privilege escalation, while the record describes code execution through custom guardrail testing.
That mismatch leaves the exact vulnerability mapping unresolved. Neither source supplies a verified fixed release for the reported key-regeneration path. Gateway operators should obtain applicable remediation from their supplier rather than infer a patch version from this report. The broader evidence of stolen identities and unauthorized billing does not depend on resolving that CVE attribution.
Update — November 19, 2026, 4:42 a.m. EST
CISA now lists CVE-2026-40217 as exploited. Its November 18 catalog addition concerns code execution through custom guardrail testing. This strengthens the exploitation assessment for that flaw; it does not resolve the key-regeneration attribution mismatch above.
Markus Vervier’s original X41 advisory, published April 8, describes an authenticated request escaping the Python restrictions at /guardrails/test_custom_code. Code runs with the proxy process’s permissions, including root in its default Docker image. The vendor advisory, published May 7, says default configurations require a proxy-admin credential to reach the endpoint. Exposure therefore depends on access to that administrative route and credential.
The vendor lists affected versions 1.81.8 through versions before 1.83.10. Its structured patch field says 1.83.10 or later, but its remediation prose says upgrade to 1.83.11 or later. Follow the explicit 1.83.11-or-later instruction for this flaw and verify the deployed artifact with the supplier. The advisory says the repair replaces the original sandbox with RestrictedPython. This historical minimum is not a recommendation to downgrade a newer maintained deployment.
If upgrading must wait, the vendor recommends blocking POST /guardrails/test_custom_code at the reverse proxy or API gateway. X41 also recommends limiting administrative access and running the proxy without root privileges. Those controls address the guardrail path; they are not a verified repair for the different key-regeneration mechanism reported by Mandiant.
Evidence status: exploitation reported by CISA, beyond X41’s controlled demonstration. The catalog gives no campaign narrative, victim count or first-attack date, and marks ransomware use Unknown. Its listed December 2 deadline belongs to the applicable federal directive, not every operator. November 18 is the catalog-addition date, not the April disclosure or a known start of exploitation.
Caveat: Google, Mandiant and X41 sell security services; the published cases provide no complete victim count or independently audited loss total. The gateway attribution mismatch remains unresolved.