New OWASP Top 10 Identifies Major AI Infrastructure Vulnerabilities

A newly released OWASP Incubator project has spotlighted often-overlooked security risks within AI infrastructure. While most AI security efforts focus on models or agent behavior, this list turns attention toward servers, data centers, and network fabrics that underpin AI systems.

AI data centers typically house thousands of GPUs, high-speed interconnects, remote management tools, shared storage fabrics, and complex schedulers. Many of these systems evolved from high-performance computing environments with trusted users. In contrast, today’s commercial GPU hosting often supports entirely separate organizations on shared hardware. Vulnerabilities once dismissed as minor in trusted environments can now expose entire customer bases to risk.

What’s New in the OWASP Top 10 Neocloud and AI Data Center Risks

The project, led by security researchers Michael Katchinskiy and Yakir Kadkoda of Lava, draws from CVEs, published incidents, vendor documentation, and academic and practitioner reports. Risks were graded based on likelihood, impact, and how readily they can be detected. Reviewers included experts from industry and academia, including Google, IBM, Dell, Nebius, and the University of Basel.

This OWASP list intentionally excludes issues like prompt injection or model misuse, since those are covered by other frameworks such as the OWASP Top 10 for LLM Applications and the MITRE ATLAS taxonomy.

The Top Infrastructure Risks Exposed

The framework ranks its biggest dangers from most to least severe. The top five—considered critical—include: 1) hardware and firmware tampering that survives resets, 2) vulnerable interconnects between GPUs, 3) weak tenant isolation or leftover data between users, 4) insecure out-of-band management tools like BMCs (baseboard management controllers), and 5) supply chain compromise affecting firmware, drivers, or OS images.

Additional risks further down the list include: flawed facility systems (like power or cooling), improper handling of model or artifact files, gaps in audits or transparency, insecure operational tools, and slow or inconsistent patching practices.

Real-World Evidence and Impacts

These risks are not hypothetical. For example, over 36,000 IPMI interfaces were found exposed online, many leaking authentication data. Some were tied to GPU or bare-metal cloud providers. Another incident involved a vulnerability (CVE-2024-54085) in AMI MegaRAC software that allowed attackers to bypass login controls. Devices made by Supermicro also remained vulnerable due to weak firmware patching. The NVIDIAScape container escape flaw (CVE-2025-23266) allowed a hostile container to escalate privileges up to root on a GPU host—matching predictions in the OWASP list.

The authors also draw attention to recent events like intellectual property theft from AI data centers, the illicit acquisition of data center designs in China, and even the theft of petabytes of data from large supercomputing centers. The pressures of GPU scarcity mean many customers compromise on security when selecting cloud providers.

Actions for Providers and Customers

Cloud providers and AI infrastructure operators can use this list as a benchmark to evaluate security maturity. Customers should demand transparency—ask whether audits include GPU hardware and whether vendors resell capacity. Be wary of providers that obscure ownership or neglect fundamentals like firmware integrity.

Concretely, defense strategies should include taking BMCs off the public internet, implementing dedicated networks for management traffic, getting firmware regularly patched, separating tenants strictly, and ensuring facilities management is robust. A 60-day checklist has been mapped out to guide buy-in and remediation.

Why this matters: As AI deployment accelerates, its foundational infrastructure becomes an attack surface that many defenders have long neglected. Compromise at the hardware, firmware, or management layer can undermine everything built on top. Cloud providers, enterprise AI teams, and cybersecurity leaders must shift scrutiny to infrastructure—not just models—to avoid high-impact, low-visibility breaches.