OWASP Incubator Project Maps Security Risks in GPU Clouds and AI Data Centers
Emerging Technology Risk · 2 sources · Updated 1d ago · First seen Nov 16, 2026

OWASP’s new Top 10 Neocloud and AI Data Center Security Risks Incubator Project addresses infrastructure weaknesses beneath AI models, applications, and agents. Its five highest-ranked risks are hardware and firmware compromise, insecure GPU interconnects, inadequate tenant isolation, insecure out-of-band management, and supply chain compromise. The framework highlights how shared GPU infrastructure adapted from trusted high-performance computing environments can expose multiple customers to attacks that conventional security monitoring may miss.
Supporting examples include internet-exposed IPMI management interfaces, the AMI MegaRAC authentication-bypass vulnerability CVE-2024-54085, and the NVIDIAScape container-escape vulnerability CVE-2025-23266. Compromised baseboard management controllers (BMCs) can provide persistent access outside normal operating-system visibility. Recommended defenses include isolating management networks, keeping BMCs off the public internet, restricting administrative access, and promptly patching firmware. Organizations buying GPU capacity should also verify whether provider audits cover GPU infrastructure and resold capacity, rather than assuming general cloud assurances extend to those environments.


