OWASP published a Top 10 list for neocloud and AI data center security so providers and customers can name infrastructure risks that sit below models, agents, and prompts. The OWASP incubator project, co-led by Lava researchers Michael Katchinskiy and Yakir Kadkoda, ranks firmware through patch-velocity failures. If you run coding agents locally but rent GPUs, store checkpoints remotely, or fine-tune on shared hardware, these are the questions your procurement checklist probably skipped.
| Question | Short answer |
|---|---|
| Am I affected? | Yes, if you use neocloud GPU APIs, bare-metal AI hosts, or third-party training clusters |
| Is there a fix? | No single patch. Provider controls plus contract requirements per risk |
| Does data leave my machine? | Model weights, datasets, and API keys can if tenant isolation or artifact handling fails |
| Does this replace OWASP LLM Top 10? | No. It stacks underneath |
| What should I ask first? | Tenant isolation, BMC exposure, and who operates resold capacity |
What the framework is (and is not)
The authoritative text lives in the GitHub repository and on owasp.org. OWASP describes it as a vendor-neutral guide to AI infrastructure: accelerators, high-speed networks, orchestration, identity, management planes, firmware, and physical facilities.
It does not cover prompt injection, insecure agent plugins, or model abuse. For those, OWASP names the OWASP Top 10 for LLM Applications, MITRE ATLAS, NIST AI RMF, and ISO/IEC 42001. Beam's MCP practical guide and agent monitoring explainer sit in that application layer. The neocloud Top 10 answers a different search intent: what breaks if the metal beneath the agent is wrong.
The authors stress that much gear comes from HPC traditions built for trusted researchers, while commercial neoclouds now host unrelated tenants on the same machines. A flaw that was tolerable in a university cluster can become catastrophic when the neighbor is a competitor's training job.
How the risks were chosen
According to the README, candidates came from security research, incident reports, CVEs, vendor documentation, and practitioner input. Reviewers included contributors from Dell, Nebius, Google, IBM, University of Basel, and others listed on GitHub. Each entry was scored on likelihood, impact, and detection difficulty, then ordered by severity.
The document also notes GPU scarcity: customers often choose whoever has capacity, not whoever is most secure, which reduces market pressure on providers and shifts diligence to buyers.
The Top 10 at a glance
| Rank | Risk | Level (published table) |
|---|---|---|
| 01 | Hardware and firmware integrity compromise | Critical |
| 02 | Network and interconnect vulnerabilities (InfiniBand, RoCE, RDMA) | Critical |
| 03 | Unsafe multi-tenant isolation and resource reuse | Critical |
| 04 | Insecure out-of-band management plane (BMC) | Critical |
| 05 | AI infrastructure supply chain compromise | Critical |
| 06 | Insecure facility and DC management systems | High |
| 07 | Insecure data and artifact handling | High |
| 08 | Certification gaps and provider transparency failures | High |
| 09 | Insecure operational infrastructure services | High |
| 10 | Vendor embargo gaps and patch velocity failures | Medium |
Each risk in the repo includes definition, impact, provider and customer mitigations, attack scenarios, and references. Below is what matters for teams that use coding agents plus rented AI compute.
Critical risks, mechanism by mechanism
01 — Firmware and hardware integrity
Compromise below the operating system can survive reinstalls, disk wipes, and tenant reassignment. Baseboard management controllers and UEFI-level persistence are the canonical example in the README. For agent developers, the lesson is: you cannot bootstrap trust from inside the guest OS alone if the board is hostile.
02 — Network and interconnect weaknesses
AI clusters depend on low-latency GPU fabrics. Misconfigured InfiniBand, RoCE, or RDMA paths can leak traffic or allow lateral movement between nodes. An agent that ships a training orchestration template with flat network assumptions can accidentally run multi-node jobs on insufficiently segmented fabrics.
03 — Multi-tenant isolation and resource reuse
Neoclouds reuse physical servers between customers. Failed cleanup can leave checkpoints, embeddings, or environment variables for the next tenant. That is especially relevant when coding agents generate Dockerfile layers or Helm charts that mount shared cache volumes by default.
04 — Insecure out-of-band management (BMC)
A compromised BMC can power-cycle hosts, flash firmware, or open remote consoles outside normal OS logging. Risk 04 cites CVE-2024-54085 on AMI MegaRAC-derived platforms. Separately, Lava research reported in 2026 found 36,872 internet-exposed IPMI hosts on UDP port 623 and 24,650 that exposed password-derived authentication material usable for offline cracking (CVE-2013-4786).
05 — Supply chain compromise
Drivers, firmware, golden images, and orchestration packages enter through many vendors. The README treats compromised artifacts as critical because they ride trusted update paths, similar to how agent users worry about typosquatted npm skills, but at the hardware and hypervisor layer. See Beam's supply-chain posts for application-layer parallels.
High risks that touch agent artifacts
07 — Insecure data and artifact handling
This is the closest overlap with daily agent work. Shared storage for datasets, prompts, logs, embeddings, checkpoints, and model weights needs access control and secure deletion. A coding agent that uploads a merged adapter or evaluation log with secrets to a provider bucket inherits whatever retention and isolation that bucket has.
Risk 10 cites CVE-2025-23266 (NVIDIAScape), a NVIDIA Container Toolkit privilege-escalation issue fixed in 1.17.8 and later, as an example of slow patch uptake widening GPU host escape risk.
08 — Certification and transparency gaps
Audits may cover software certifications while GPU rows, resold bare metal, or subcontracted operators sit outside scope. Ask whether the provider resells another operator's capacity and whether your contract can require answers aligned with risk 08.
09 — Operational infrastructure services
Schedulers, provisioning APIs, and automation consoles that can push workloads fleet-wide are high-value targets. Compromise there is comparable to compromising the CI/CD plane that your coding agent already touches, but at cluster scale.
10 — Patch velocity and embargoes
Slow or disruptive patching lengthens n-day windows. That connects to offensive automation described in Google GTIG's patch-to-exploit reporting and GTIG vulnerability trend data: infrastructure CVEs on management planes deserve the same urgency as app CVEs your agent deploys.
Procurement questions to paste into a provider review
The OWASP project is written for contract language as much as for architecture reviews. When a coding agent has already drafted your Kubernetes manifests, these human questions still need answers from the operator:
- Who physically racked the GPUs, and whether supply-chain attestation covers firmware on those boards (risk 01 and 05).
- Whether east-west RDMA traffic is segmented per tenant or only isolated at the hypervisor (risk 02).
- What happens to VRAM, local SSD, and memory-backed caches when your job ends and another customer's job starts on the same host (risk 03).
- Whether BMC or IPMI addresses are reachable from the public internet, and whether your contract allows independent verification (risk 04).
- Which orchestration API version you are pinned to, and how quickly the provider patches known management-plane CVEs compared with guest OS images (risk 10).
Keep answers in writing. Agent-generated architecture diagrams are not evidence of isolation.
Threat actors the framework lists
The README's threat model is deliberately broad: external attackers, compromised cloud customers, malicious insiders, suppliers, in-transit hardware tampering, and "AI agents acting on behalf of any of the above." The last item matters for agent teams: a compromised coding agent credential is classified like a malicious customer workload in a shared cluster.
Enterprise compromise paths (phishing, MFA gaps, CI/CD theft) are called cross-cutting initial access before infrastructure-specific failures. That matches how Storm-0940 and ARTEX cases began with identity, then reached high-value systems.
What this means if you run AI coding agents
Most readers are not neocloud operators. They are developers whose agents provision cloud resources, write training configs, or store API keys in cluster secrets. Practical checks:
-
Map where compute runs. If the agent only edits local files, risks 01–06 are mostly your laptop vendor's problem. If you
kubectl applyor call a neocloud API, you inherit risks 03, 07, and 09. -
Ask providers the README's customer questions. Does their audit cover GPU hardware? Do they resell capacity? Is the BMC network internet reachable? Lava's project page repeats these for procurement.
-
Treat checkpoints like secrets. Version-pin containers, encrypt weights at rest, and avoid shared cache paths in agent-generated job YAML. Qwen cyber-training filters protect model behavior, not disk neighbors.
-
Separate agent keys per environment. A single API key embedded in an agent skill and copied to a cluster secret is both an LLMjacking target and a tenant artifact leak risk. See GTIG on stolen model keys.
-
Do not confuse MCP scanning with data-center assurance. Beam's 11-pattern skill and MCP scan catches dangerous instructions before you run a skill locally. It does not attest that a rented A100 host wiped the previous tenant's VRAM-resident state.
-
Combine frameworks. Use OWASP LLM Top 10 for agent behavior, this list for where the agent's compute lives, and your attacks timeline discipline for exploited CVEs on orchestration tools (November 2026 timeline).
What Beam can and cannot do here
Beam is local-first observability and heuristic scanning for coding agents on your machine: skill and MCP config patterns, optional hook deny for Claude Code and Codex, redaction before event storage. It does not assess neocloud BMC policies, GPU fabric segmentation, or provider firmware practices.
Use Beam when agents generate infrastructure-as-code or touch secrets on the laptop. Use the OWASP neocloud Top 10 when you choose or renew the provider that runs the resulting workloads.
Frequently asked questions
What is the OWASP Top 10 for neocloud and AI data centers?
It is an OWASP incubator framework listing the ten most consequential security risks in neoclouds and AI data centers, from firmware integrity through patch velocity. The GitHub README and owasp.org project page publish definitions, attack scenarios, and mitigations for providers and customers.
Does this list cover prompt injection or MCP skills?
No. The project scope is infrastructure under the model. OWASP points readers to the OWASP Top 10 for LLM Applications, MITRE ATLAS, NIST AI RMF, and ISO/IEC 42001 for application-layer and agent-behavior risks.
What is the difference between OWASP Top 10 LLM and neocloud Top 10?
The LLM Top 10 addresses model and application misuse such as prompt injection and insecure output handling. The neocloud Top 10 addresses GPUs, BMCs, RDMA fabrics, multi-tenant hardware reuse, and facility systems that host those workloads.
Why should someone running Claude Code or Codex on a rented GPU care?
Many teams fine-tune, run evals, or host open-weight models on neocloud GPUs while coding agents push configs and API keys. Weak tenant isolation or leaky checkpoints can expose weights, datasets, or secrets even when your agent harness is configured safely on the laptop.
Which risks are rated critical?
Risks 01 through 05 are critical in the published table: hardware and firmware integrity, network and interconnect weaknesses, unsafe multi-tenant isolation, insecure out-of-band management (BMC), and AI infrastructure supply chain compromise.
What real incidents does the framework cite?
The project README and risk documents cite public reporting on US AI data center IP theft, a conviction for stealing AI data center designs, claimed Tianjin supercomputing data exfiltration, IPMI exposure research, AMI MegaRAC CVE-2024-54085 in Risk 04, and NVIDIA Container Toolkit CVE-2025-23266 in Risk 10.
How should I use the list in procurement?
Ask whether security audits cover GPU hardware and resold capacity, require BMC networks to stay off the public internet, document tenant cleanup between jobs, and map which risks the provider mitigates versus delegates to you.
Can Beam replace neocloud infrastructure controls?
No. Beam scans skills and MCP configs and watches local agent actions on your machine. It does not validate a provider's BMC hardening or GPU isolation. Use it for the workstation agent layer; use the OWASP list when you choose where that agent's compute and artifacts live.
Related reading
- BIS remote-access chip export controls for neocloud agents
- Trump overseas AI data center tax and agent residency inventories
- MCP security practical guide
- What is AI agent monitoring
- GTIG exploit_agent and patch windows
- LLMjacking and stolen keys
- November 2026 attacks timeline
OWASP project text, Lava's November 2026 announcement, and CVE references were reviewed on November 16, 2026 UTC. Provider implementations vary; confirm controls in contract and audit artifacts, not marketing pages alone.

