Skip to content
RESEARCH INDEX BREACHROAD / INTELLIGENCE NOTE

LLMjacking: how criminals monetise cloud AI access

Google documents LLMjacking campaigns where stolen cloud credentials generated more than $100,000 in daily charges. Learn how to detect and stop it.

PUBLIC RESEARCH
AUTHOR
/ CEO of Breachroad · OSCP · PNPT
PUBLISHED
7 November 2026
READING TIME
18 min read
TOPIC
AI Security
LLMjacking: how criminals monetise cloud AI access

LLMjacking is the theft of access to paid language models and its use to generate output at the victim’s expense. Google Threat Intelligence Group described incidents on 2 November 2026 involving multiple cloud and model providers, organisations in several sectors and charges above $100,000 in a single day. In some environments, cumulative unauthorised usage exceeded one million dollars.

The factual source is Google’s “LLMjacking: Monetizing compromised cloud models”, which covers activity Mandiant observed during 2026. This is not a theoretical prompt-injection scenario. An operator obtains valid cloud credentials or a session, discovers available models and then consumes or resells their inference capacity.

The criminal business model

An operator does not need to steal training data or compromise the model itself. The ability to send requests billed to an organisation is enough. Stolen access is fed into intermediary services that combine multiple accounts and models behind one interface. Google calls these LLM Mixers: aggregators distribute load, obscure its origin and sell cheap or nominally unlimited access.

The economics resemble cryptojacking, except the unit of value is a token and access to a model. High quotas, premium models, long contexts and repeated generations can increase the bill very quickly. Because legitimate AI workloads can be bursty, a basic “many requests” alert may not be sufficient.

Confirmed initial-access paths

Google identifies public-facing services, compromised IAM identities and service accounts, GitHub tokens and active session tokens among initial-access classes. The common factor is a credential that can enumerate models or submit inference requests. Attackers automate discovery and rapidly test limits.

This distinction matters: the incident starts in the identity, cloud and billing layers, not in a prompt. Prompt-injection defences do not stop an API client holding a valid key. That is why a secure AI implementation programme must include IAM, secrets, budgets and telemetry rather than content filters alone.

Signals worth correlating

The strongest alert joins identity evidence with model-usage evidence. Higher spending alone may reflect a product launch. A new source address may belong to an employee travelling. Combined with model enumeration, a region change and high-volume inference outside the established profile, those signals form a much stronger hypothesis.

At minimum, monitor:

  • first model use by an identity;
  • calls from a new region, ASN or client tool;
  • enumeration of models and quotas without a matching application change;
  • a sudden rise in input and output tokens;
  • use of a premium model instead of the approved version;
  • many identical or near-identical requests in a short period;
  • continuous activity from an account previously tied to one pipeline;
  • quota errors, attempts to raise quotas and creation of new keys;
  • changes to budgets, billing projects or log destinations.

Keep the identity, model, region, project, token counts, status, latency and a privacy-compatible content fingerprint. Full prompts may contain sensitive information, so security visibility should not silently become an uncontrolled content archive.

Responding to detection

Stop the cost first, while preserving evidence. Disable or restrict the compromised credential, force session termination, block sources through policy and reduce quota for unused models. At the same time, export IAM, audit, inference, CI/CD, repository and billing records to a location the intruder cannot alter.

Then establish the access path. If a key appeared in a public repository, rotation alone is insufficient: remove it from history and examine forks, build artefacts and CI logs. For a stolen session token, investigate the originating endpoint and authentication flow. For a service account, review all roles, impersonation paths and secrets reachable in the same project.

Notify the cloud owner and finance team. A provider may help with records, limits and charge review, but the organisation needs its own timeline. Do not assume a refund. Budgets and alerts are safety devices, not substitutes for access control.

LLMjacking-resistant architecture

The first rule is to avoid long-lived keys in source code and user devices. Applications should receive short-lived workload identity, while a backend broker enforces the allowed model, region, maximum context and limits. End users rarely need a provider key directly.

The second rule is minimum scope. Separate projects for production, testing and experiments reduce blast radius. An application account should not be able to increase its own quotas, disable logs or create more keys. High administrative privilege should remain separate from permission to run inference.

The third rule is cost control at several levels: per request, user, service, model and time window. Hard limits can affect availability, so critical processes need a designed degraded mode. It is better to switch deliberately to a cheaper model or a queue than to discover the incident from an invoice.

Our guide to non-human identities for AI agents expands the IAM problem, while AI agent sandbox architecture explains how to isolate a model from tools and secrets. The two controls must work together: a sandbox holding an overpowered token remains dangerous.

Threat modelling for an AI team

Draw the flow from user to model and billing system. Where is identity created? Who selects the model? Where is the secret stored? Who can raise a limit? How is an application request correlated with the provider record? How quickly can access be revoked without stopping every service?

A tabletop exercise can begin with a financial alert on Saturday. Participants then receive new facts: a new ASN, a service account also used by CI, a token found in an old artefact and accelerating charges. The aim is not to guess an actor but to make containment, continuity, evidence and communication decisions.

What LLMjacking is not

Prompt injection changes model behaviour inside a legitimate session. Model extraction attempts to reconstruct model properties. Data theft targets the confidentiality of inputs or systems. LLMjacking primarily monetises someone else’s permission and budget, although a single incident may combine these classes. Precise naming directs ownership correctly: IAM and FinOps matter as much as the AI team.

Not every large bill is an intrusion. A loop bug, runaway agent, missing cache or tariff change can produce a similar symptom. The investigation must test both hypotheses. Normal usage baselines and deployment records help distinguish intrusion from a cost failure.

Source facts and Breachroad analysis

Source facts: Mandiant observed 2026 incidents across regions and industries, access through compromised credentials, and services reselling it as LLM Mixers. Documented cases exceeded $100,000 of daily charges and one million dollars cumulatively. Breachroad analysis: correlation rules, layered limits, project separation and the tabletop scenario follow from the threat model and are not claims about every incident Google investigated.

What to do this week

Search for model-provider keys in repositories, CI secrets and container images. Map identities permitted to invoke models. Configure anomaly-based cost alerts and hard limits where interruption is safe. Verify that the on-call responder can revoke a token within minutes and restore service with a clean identity.

Finally, run a controlled architecture review. AI red teaming can test whether an agent or application can be made to reveal a secret or select an unapproved model. Cybersecurity and secure-AI training aligns ML, cloud, SOC and FinOps teams around one response plan. The strongest LLMjacking defence makes a stolen credential short-lived, narrow and immediately visible in cost telemetry.

FinOps as a security sensor

Budgets need informational, warning and emergency thresholds. Informational creates a contextual ticket; warning restricts a premium model or concurrency; emergency denies new work outside approved services. Automatic blocking must account for critical processes so an attacker cannot deliberately exhaust quota to cause denial of service.

The cost model should know expected spend by user, function, product and hour. Alert on rate deviation rather than only a monthly total. Because model prices differ, one hundred similar requests can have radically different financial impact.

Secrets across the software lifecycle

Repository scanning should include history, pull requests, comments, CI artefacts, logs and container images. A key removed from the latest commit may remain valid in a fork or cache. Revoke first, then remove the material and investigate usage. Reversing that order extends exposure.

Prefer workload-identity federation between CI and cloud. A pipeline receives a short token for one project and cannot raise quotas. Production identity should not be available to unapproved branches or forked pull requests.

Analysing content without losing privacy

Detecting repeated resale traffic can use feature summaries: model, token count, language, length, embedding or a private fingerprint. Full conversation storage is not always necessary. Content access should require an incident need, be audited and time-limited.

If complete prompts are essential, use a dedicated evidence repository and redaction policy. Finance needs cost and project, not customer content. Separating views reduces secondary exposure during an LLMjacking investigation.

Testing revocation

Quarterly, create a test key and simulate its appearance in a controlled repository. Measure detection, revocation, quota restriction, copy discovery and service restoration with a clean identity. Test every region and API-gateway cache.

Confirm that the old token actually fails and alerts identify an owner. An AI security audit can compare policy with effective roles and limits.

Maturity metrics

Report long-lived keys, ownerless identities, revocation time, inference-log coverage, anomalies found before invoicing and maximum possible hourly cost. A lower average bill is insufficient; the cap must hold in the most expensive allowed scenario.

Safe incident-closure criteria

After LLMjacking, confirm revocation of every credential copy, closure of the leak path, usage-history review, termination of unauthorised spend and restoration with a clean identity. Check that the intruder did not create another key or role. FinOps should reconcile billing with technical logs, while legal assesses whether prompts contained reportable data.

After one month, repeat secret scanning and a revocation drill. If cost can rise faster than an alert reaches on-call staff, lower the limit or add an automatic safety control.

SHARE / COPY