A new OWASP Incubator Project shows a missed area in AI security: servers, networks, and data centers for AI systems get much less focus than models, apps, and agents. The OWASP Top 10 Neocloud and AI Data Center Security Risks aims to address this issue. It provides neocloud providers, their clients, and AI data center operators a common list of the major infrastructure risks.
AI data centers hold many GPUs, fast networks, remote controls, shared storage, and scheduling tools all in one location. Many of these tools came from high-performance computing, where users were trusted. Now, the same parts are used in commercial GPU clouds that different companies use together. A small flaw that was not a big deal before can now endanger many customers.
Many security tools can’t detect this layer. They look at operating systems, apps, and regular networks. They frequently overlook firmware, management ports, and unique links that connect GPU nodes. If the base is hacked, nothing above can be trusted.
Michael Katchinskiy and Yakir Kadkoda of the security firm Lava lead the project. The team selected risks from security research, incident reports, published CVEs, vendor documents, and practitioner talks.
Each candidate was scored on likelihood, impact, and detectability. Reviewers came from companies and schools such as Dell, Nebius, Google, IBM, and the University of Basel.
OWASP Top 10 Exposes Neocloud AI Risks
The project does not deal with prompt injection or model misuse. Other guides, like the OWASP Top 10 for LLM Applications and MITRE ATLAS, take care of those issues. OWASP lists the risks from highest to lowest danger. The top five are called critical.
These include hardware and firmware changes, weak links between GPUs, bad separation between users, unsafe remote management, and supply chain issues.
Weak facility systems, careless data and model file handling, gaps in certifications and provider honesty, unsafe operation tools, and slow patching are lower on the list.
Firmware and management controllers need attention. A baseboard management controller, or BMC, allows admins to manage a server even when the operating system is down. If a hacker gains control of it, the access can remain after a reinstall and can move to the next customer with the server.
OWASP Top 10 Risks:
| # | Risk | What can go wrong |
|---|---|---|
| 1 | Hardware and firmware integrity compromise | Hidden tampering below the OS survives reinstalls and resets, so access persists after a server is reassigned |
| 2 | Network and interconnect vulnerabilities | Weakly secured InfiniBand, RoCE, and RDMA links allow unauthorized access, data interception or changes, and movement between GPU nodes |
| 3 | Unsafe multi-tenant isolation and resource reuse | Failed isolation or uncleaned servers expose one tenant’s data, workloads, or infrastructure to another |
| 4 | Insecure out-of-band management plane | A hacked BMC can restart servers, alter firmware, open remote consoles, or take over whole fleets |
| 5 | AI infrastructure supply chain compromise | Compromised hardware, firmware, drivers, OS images, or updates bring in malicious code or hidden access |
| 6 | Insecure facility and data center management systems | Weak cooling, power, and access control systems can disrupt operations or leak sensitive infrastructure details |
| 7 | Insecure data and artifact handling | Poor access controls or cleanup lead to stolen model weights, tampered training data, or recovered leftover files |
| 8 | Certification gaps and provider transparency failures | Audits may skip the GPU layer, and undisclosed resale of capacity hides who runs the hardware |
| 9 | Insecure operational infrastructure services | Compromised deployment and management tools let attackers push malicious workloads, change configs, or disable controls |
| 10 | Vendor embargo gaps and patch velocity failures | Late disclosure and slow or disruptive patching give attackers a longer window to exploit known flaws |
This is not only a theory. Researchers found 36,872 IPMI interfaces online, and around 67 percent leaked password data before users logged in.
Some were from GPU and bare-metal cloud providers. Earlier, a problem in AMI MegaRAC software (CVE-2024-54085) allowed attackers to bypass login on affected servers. Our report on BMC firmware flaws showed that weak Supermicro fixes left gaps. Tenant isolation is another weak area.
The NVIDIAScape container flaw, CVE-2025-23266, allowed a bad container to get root access on a GPU host. It affected toolkit versions up to 1.17.7. This is the type of issue that the OWASP list warns against.
Related News:
OWASP Releases Security Checklist for Generative AI Deployment
OWASP Unveils GenAI LLM Top 10 2026 For Modern AI APPS
OWASP Top-10 2025: Revised Version Include Two New Categories
InfoSecBulletin Cybersecurity for mankind
