The infrastructure behind artificial intelligence is becoming more specialized, interconnected, and distributed. While much of the attention surrounding AI security focuses on models, applications, and data, the underlying computing environments introduce another set of vulnerabilities that organizations must consider.
AI data centers rely on extensive combinations of GPUs, networking equipment, firmware, management controllers, storage systems, and orchestration technologies. These components must operate together, often within environments where multiple organizations share physical infrastructure.
That creates a security challenge that extends beyond traditional software vulnerabilities. Lava Security is addressing the issue through its Top 10 Neocloud and AI Data Center Security Risks framework, which has become an official OWASP project.
The initiative, developed by Yakir Kadkoda, CTO and Co-Founder of Lava Security, and Michael Katchinskiy, Head of Security Research, aims to establish a shared reference for identifying security weaknesses across the infrastructure supporting AI workloads.
Why Infrastructure-Level Vulnerabilities Matter
Many technologies used in modern AI data centers were originally developed for trusted high-performance computing environments. Their adoption within commercial neocloud environments changes the security considerations, particularly when unrelated organizations operate on shared physical infrastructure.
Bare-metal cloud deployments provide one example. Customers may receive root access to rented servers, allowing them to control their computing environments. However, unrestricted access to management interfaces can also create opportunities to modify hardware boot configurations.
Those changes may survive an operating system reinstall. If the configurations are not reset before the server is reassigned, a subsequent customer could inherit a compromised environment.
This illustrates a broader concern: security controls must account for hardware and firmware states, management interfaces, and the processes used to transfer infrastructure between customers.
“AI infrastructure introduces attack surfaces that traditional security approaches may not adequately address. Organizations need to understand what happens below the operating system, particularly when hardware, firmware, and management interfaces are shared across customers,” said Michael Katchinskiy, Head of Security Research at Lava Security. “This framework helps make those risks easier to identify and provides practical guidance for addressing them.”
Ten Areas of Risk Across the AI Infrastructure Stack
The OWASP Top 10 Neocloud and AI Data Center Security Risks examines vulnerabilities across hardware, firmware, networking, management, and operational systems.
The framework includes risks associated with compromised server firmware and weaknesses in high-speed networking technologies such as InfiniBand and RDMA. It also addresses inadequate tenant isolation and insecure out-of-band management interfaces.
Beyond these technical components, the research examines supply chain threats, data center facility systems, shared storage and AI artifacts, incomplete security certifications, centralized operational services, and delays in vulnerability patching.
The framework takes an infrastructure-wide approach rather than examining individual technologies in isolation. This matters because interconnected systems can create broader exposure when weaknesses in one component affect other parts of the environment.
Each risk category provides potential attack scenarios, real-world implications, and practical mitigation recommendations. Infrastructure providers and operators can use these insights to assess their security practices, identify gaps, and prioritize improvements.
Organizations purchasing GPU cloud capacity can also reference the framework when evaluating the security controls protecting their workloads.
Research Developed With Industry Input
Lava Security initially published the framework in July 2026. Its development drew on disclosed vulnerabilities, incident reports, technical documentation, and the company’s research into AI infrastructure security.
The initial work also incorporated feedback from more than 20 security and infrastructure practitioners representing neocloud providers, hardware original equipment manufacturers, security companies, universities, and research institutions.
That participation reflects the range of expertise required to understand infrastructure security. Hardware vendors, infrastructure engineers, operators, and independent researchers each encounter different aspects of the same environment.
For Kadkoda, establishing the initiative within OWASP provides a way to bring those perspectives together.
“Securing AI infrastructure requires looking at the entire environment, not just individual components. Hardware, networking, management systems, and workloads are interconnected, and risks in one area can affect the security of the broader infrastructure,” said Yakir Kadkoda, CTO and Co-Founder of Lava Security. “Developing this framework within OWASP gives the community an opportunity to contribute its expertise and continuously improve how these risks are understood and addressed.”
Keeping Pace With AI Infrastructure Development
The project will continue evolving through an open development process on GitHub. Security researchers, infrastructure engineers, hardware vendors, and operators can review its contents, challenge assumptions, propose revisions, and contribute additional research or mitigation recommendations.
Contributors will be credited in future releases.
The framework is designed to develop alongside changes in AI infrastructure, including new GPU architectures, networking technologies, management systems, and deployment models.
Those interested in exploring the research can visit the official OWASP project page or access the GitHub repository to participate in its development.
As AI infrastructure continues to expand, establishing a common understanding of the vulnerabilities affecting its underlying systems could help providers and users approach security with a more comprehensive view of the environments supporting their workloads.


