Cloud GPU hijacking has historically been associated with cryptocurrency mining, but attackers are increasingly finding other uses for compromised computing resources. From unauthorized AI inference to model training and automated offensive tooling, cloud environments are becoming targets for attackers interested in stealing computational capacity rather than simply deploying cryptominers.
Graphics processing units (GPUs) have become central to artificial intelligence infrastructure, supporting machine learning training, large language model (LLM) inference, and high-performance computing. Cloud providers offer access to this hardware through GPU-backed virtual machines, managed AI platforms, and containerized infrastructure, allowing organizations to provision substantial computing capacity without maintaining dedicated hardware.
The same accessibility creates opportunities for attackers who obtain cloud credentials or compromise exposed AI services. Rather than purchasing their own infrastructure, threat actors can redirect existing resources, provision expensive GPU instances, or consume paid inference services through compromised accounts. The financial impact can be substantial, but unauthorized resource consumption can also introduce persistence mechanisms, expose sensitive information, and provide infrastructure for subsequent attacks.
How Attackers Gain Control of Cloud GPU Infrastructure
Cloud GPU hijacking frequently begins with compromised identities, exposed management interfaces, or vulnerable applications running inside cloud environments. The attacker does not necessarily need to exploit a GPU driver or compromise the physical hardware. Access to the management plane or the application controlling GPU resources may be sufficient to redirect computing capacity.
Stolen cloud credentials are particularly valuable. An attacker who obtains an AWS Identity and Access Management (IAM) access key with excessive permissions may be able to enumerate existing resources, launch EC2 instances, modify security groups, and interact with Amazon Bedrock. Similar risks exist across other cloud providers, where compromised service accounts can authorize unauthorized compute deployments or access managed AI services.
In February 2026, Sysdig published an investigation detailing an AWS intrusion in which an attacker obtained credentials from publicly accessible S3 buckets containing retrieval-augmented generation (RAG) data. The attacker used Lambda function code injection to escalate privileges, moved across 19 AWS principals, and accessed services including Bedrock and EC2.
After obtaining elevated permissions, the attacker attempted to provision an AWS p5.48xlarge instance before successfully launching a p4d.24xlarge instance equipped with eight NVIDIA A100 GPUs. The attacker prepared scripts that referenced CUDA, PyTorch, DeepSpeed, and model training workflows, along with a publicly accessible JupyterLab interface. Sysdig reported that the instance was terminated after approximately five minutes, leaving the attacker’s intended use of the GPU resources uncertain.
This incident demonstrates how cloud identity compromise can become an infrastructure abuse problem. The attacker moved from credentials stored in an accessible bucket to administrative permissions and expensive GPU-backed resources within the same environment. Traditional access controls intended for application development became part of the attack path.
Exposed AI orchestration platforms present a separate entry point. Frameworks such as Ray allow distributed workloads to execute across multiple compute nodes, often with access to GPUs and sensitive datasets. If their administrative interfaces are exposed without appropriate access restrictions, attackers may be able to submit jobs directly to the cluster.
Oligo Security documented this behavior in its ShadowRay 2.0 investigation. The campaign exploited exposed Ray Jobs APIs associated with CVE-2023-48022, allowing unauthorized command execution across affected clusters. Attackers leveraged Ray’s scheduling functionality to distribute malicious workloads, establish persistence, and deploy cryptocurrency miners. Researchers also documented attempts to conceal resource consumption and access sensitive information.
The vulnerability remains disputed by Ray’s maintainers, who identify the underlying behavior as a feature intended for trusted environments. The distinction does little to change the operational risk of exposing an administrative interface capable of executing arbitrary workloads.
From Cryptomining to Unauthorized AI Inference
Cryptojacking remains a common objective for attackers targeting computing infrastructure. Mining software can consume available CPU and GPU resources, generating cryptocurrency revenue at the expense of the compromised organization. The resulting symptoms may include elevated utilization, reduced application performance, unexpected network connections, and increased infrastructure costs.
AI services introduce another form of resource abuse known as LLMjacking. Originally documented by Sysdig in 2024, LLMjacking involves obtaining unauthorized access to paid language model services or self-hosted inference infrastructure. Rather than relying on cryptocurrency mining, attackers consume computational resources through model requests, resell access to compromised services, or incorporate stolen inference capacity into their own applications.
The distinction between GPU hijacking and LLMjacking matters. GPU instance hijacking involves control over infrastructure capable of running arbitrary workloads. LLMjacking can occur through stolen API credentials without giving an attacker direct access to the underlying hardware. Both activities exploit computing resources that belong to another organization, but their technical mechanisms and detection requirements differ.
In a November 2026 report, Google Threat Intelligence Group and Mandiant described multiple incidents involving compromised enterprise access to cloud-hosted AI models. Investigators observed fraudulent resource consumption exceeding $100,000 in a single day, with cumulative charges reaching $1 million at some affected organizations before access was revoked.
The attackers used exposed credentials, compromised service accounts, and stolen session tokens to access cloud environments. Some enabled previously unused AI services, generated additional API keys, and subscribed compromised accounts to pay-as-you-go offerings. Others deployed proxy infrastructure that could consolidate access to multiple AI providers through a single interface.
Researchers also identified an emerging market for what they called LLM Mixers, services that aggregate unauthorized model access and resell it through proxy gateways. These systems allow operators to distribute requests across compromised accounts, manage available capacity, and conceal the original sources of their inference services.
This changes the economics of cloud resource theft. A compromised account becomes a source of billable API requests that can be redistributed to other users. The operator may never need to maintain a cryptominer or interact directly with a GPU, yet the victim still absorbs the costs.
Stolen AI Compute Is Also Being Used for Offensive Tooling
Financial theft is one part of the problem. Researchers have also documented attackers using compromised AI infrastructure to support the development of offensive security tools.
In June 2026, Sysdig observed an operator abusing an exposed Ollama server as the inference backend for an automated vulnerability assessment and penetration testing framework.
The tool, identified as VAPT, used a staged workflow to analyze network services, associate software versions with known vulnerabilities, evaluate web application behavior, construct proof-of-concept payloads, and plan subsequent exploitation activity. Its orchestration logic allowed model responses to inform the next operation without requiring the operator to manually perform each analytical step.
The server was accessible without authentication on Ollama’s default TCP port, 11434. Rather than invoking a commercial model API, the operator directed requests to an externally accessible inference endpoint. This provided model access without paying for the underlying computation.
Sysdig observed the framework being developed and tested against private laboratory environments, rather than confirming its deployment against unrelated public targets. The activity still demonstrated that exposed inference capacity could support the development of automated exploitation tooling.
The implications extend beyond unauthorized model queries. An attacker with access to substantial inference capacity can use it for repeated code analysis, vulnerability triage, credential analysis, and other computationally demanding operations. The compromised infrastructure effectively becomes part of the attacker’s development environment.
This creates a different detection problem from traditional cryptojacking. The malicious activity might consist entirely of requests to an otherwise legitimate model server. There may be no cryptocurrency mining pool connection, recognizable malware executable, or unusual system process associated with the requests.
AI Supply Chains Introduce Additional GPU Hijacking Risks
Compromised cloud identities are not the only way attackers can gain access to GPU infrastructure. Machine learning pipelines frequently retrieve models, dependencies, and container images from external repositories. These components can introduce executable code into environments with access to expensive compute resources and sensitive data.
In October 2026, Pluto Security documented a malware campaign involving Hugging Face model repositories. Researchers identified 31 malicious repositories distributed across multiple accounts. The models impersonated small Qwen2 checkpoints and targeted GPU environments running self-hosted inference services.
The campaign abused the Hugging Face Transformers option trust_remote_code=True, which permits Python code supplied by a model repository to execute during model initialization. The malicious repositories used this functionality to install cryptocurrency miners and, in certain variants, establish reverse-shell access.
The attack did not require exploiting the GPU itself. The execution path originated from the model-loading process, where an application trusted executable code obtained from an external repository.
This matters for organizations that routinely download third-party models into training or inference environments. A GPU server may also hold API tokens, cloud credentials, object storage permissions, proprietary datasets, and model checkpoints. Once an attacker obtains code execution within that environment, those resources may become accessible depending on the workload’s permissions and isolation controls.
A compromised model repository can therefore create consequences beyond unauthorized GPU utilization. Persistent access, credential theft, and exposure of proprietary AI assets become additional risks, particularly when training environments operate with broad access to cloud storage or service accounts.
Why Traditional Monitoring May Miss GPU Hijacking
GPU-intensive activity is expected in environments running machine learning workloads. Training jobs can legitimately consume nearly all available GPU memory and compute capacity, sometimes for extended periods. High GPU utilization alone is consequently an unreliable indicator of compromise.
Attackers can also distribute activity across accounts, regions, and workloads. Short-lived virtual machines, temporary service accounts, and proxy infrastructure complicate investigations that rely exclusively on individual host activity or familiar malicious IP addresses.
Effective detection requires correlating cloud control-plane activity with workload behavior and application-level telemetry.
For AWS environments, CloudTrail can provide visibility into suspicious operations such as RunInstances, CreateAccessKey, AttachUserPolicy, and UpdateFunctionCode. Investigators should examine resource provisioning involving unexpected GPU instance families, unusual instance launches across regions, changes to IAM permissions, and modifications to security groups.
Where Amazon Bedrock is involved, model discovery and invocation activity deserves separate attention. Operations such as ListFoundationModels, ListInferenceProfiles, InvokeModel, and Converse can help identify suspicious activity when the relevant audit logging is configured. Monitoring should account for unexpected model access, abrupt changes in request volume, and identities invoking models outside their normal usage patterns.
For Google Cloud, relevant activity may include unauthorized Compute Engine instance creation, service account key generation, unexpected AI API enablement, and Vertex AI inference operations such as PredictionService.GenerateContent. Certain inference activity requires Data Access audit logging to be explicitly enabled, creating potential visibility gaps in environments relying on default settings.
At the workload level, organizations should correlate GPU telemetry, container activity, process execution, and outbound network connections. NVIDIA Data Center GPU Manager (DCGM) metrics, Kubernetes audit events, container runtime telemetry, and inference server access logs can provide useful context.
For example, a GPU-backed container that begins running an unexpected executable, establishes outbound connections to unfamiliar infrastructure, and consumes GPU resources outside scheduled workloads represents a substantially stronger detection opportunity than high utilization alone.
The same principle applies to self-hosted inference servers. Unfamiliar client IP addresses, unusual model-loading requests, unexpected model registrations, and sustained inference activity without an associated application workload should be investigated.
Reducing the Risk of Unauthorized Compute Usage
Defending against cloud GPU hijacking requires controls across cloud identity management, AI infrastructure configuration, and runtime monitoring.
Organizations should begin by limiting permissions associated with service accounts, CI/CD pipelines, development environments, and model-serving applications. Access to GPU instance provisioning, service quota modifications, IAM administration, and paid AI model services should be granted according to operational requirements rather than broadly available cloud roles.
Administrative interfaces for Ray, Ollama, JupyterLab, and similar platforms should remain inaccessible from untrusted networks. Where remote access is required, authenticated gateways, private networking, and restrictive firewall rules can prevent unauthorized users from submitting workloads or interacting with model endpoints.
Model dependencies also require scrutiny. Organizations should review executable code distributed with third-party models, restrict untrusted model-loading behavior, and apply isolation controls to workloads that require external components. Secrets and sensitive datasets should not be unnecessarily accessible to model-serving processes.
Detection teams should establish baselines for expected GPU instance types, authorized regions, inference models, service accounts, and normal API consumption. Budget notifications provide an additional warning mechanism, but they should not be treated as automatic spending limits. Provider-supported service quotas, instance restrictions, and model-specific access controls can limit the resources available to a compromised identity.
Incident response procedures should account for more than stopping the unauthorized workload. Investigators need to identify the initial credential exposure, revoke compromised sessions and API keys, review IAM modifications, locate persistence mechanisms, and determine whether the attacker accessed data or deployed additional infrastructure.
How Can Netizen Help?
Founded in 2013, Netizen is an award-winning technology firm that develops and leverages cutting-edge solutions to create a more secure, integrated, and automated digital environment for government, defense, and commercial clients worldwide. Our innovative solutions transform complex cybersecurity and technology challenges into strategic advantages by delivering mission-critical capabilities that safeguard and optimize clients’ digital infrastructure. One example of this is our popular “CISO-as-a-Service” offering that enables organizations of any size to access executive level cybersecurity expertise at a fraction of the cost of hiring internally.
Netizen also operates a state-of-the-art 24x7x365 Security Operations Center (SOC) that delivers comprehensive cybersecurity monitoring solutions for defense, government, and commercial clients. Our service portfolio includes cybersecurity assessments and advisory, hosted SIEM and EDR/XDR solutions, software assurance, penetration testing, cybersecurity engineering, and compliance audit support. We specialize in serving organizations that operate within some of the world’s most highly sensitive and tightly regulated environments where unwavering security, strict compliance, technical excellence, and operational maturity are non-negotiable requirements. Our proven track record in these domains positions us as the premier trusted partner for organizations where technology reliability and security cannot be compromised.
Netizen holds ISO 27001, ISO 9001, ISO 20000-1, and CMMI Level III SVC registrations demonstrating the maturity of our operations. We are a proud Service-Disabled Veteran-Owned Small Business (SDVOSB) certified by U.S. Small Business Administration (SBA) that has been named multiple times to the Inc. 5000 and Vet 100 lists of the most successful and fastest-growing private companies in the nation. Netizen has also been named a national “Best Workplace” by Inc. Magazine, a multiple awardee of the U.S. Department of Labor HIRE Vets Platinum Medallion for veteran hiring and retention, the Lehigh Valley Business of the Year and Veteran-Owned Business of the Year, and the recipient of dozens of other awards and accolades for innovation, community support, working environment, and growth.
Looking for expert guidance to secure, automate, and streamline your IT infrastructure and operations? Start the conversation today.














You must be logged in to post a comment.