AI fabric security is not a bolt-on — it is a set of mechanisms that operate at every layer from NIC ASIC to fabric management. The key practitioner takeaways:
- iptables protects nothing on an RDMA fabric. The kernel network stack is not in the RDMA data path. GID filtering, PKeys, and RKEY entropy are the actual enforcement points.
- GID filtering (ROCE_ADDR_FILTER_ENABLE=1) should be the default on every ConnectX-7 in a multi-tenant cluster. The factory default is disabled. Fix this at provisioning time.
- Low-entropy RKEYs are a real production vulnerability, not a theoretical one. Check kernel ASLR settings and verify RKEY values on freshly deployed nodes.
- Spectrum-X isolation requires both VRF routing and GBP. VRFs prevent inter-tenant routing; GBP prevents intra-VRF lateral movement. Neither alone is sufficient.
- PKey partitioning in InfiniBand is hardware-enforced — the most reliable isolation available. The Limited membership model for shared storage nodes is the correct pattern for multi-tenant IB clusters.
- BlueField-3 (B200 only) provides an independent enforcement plane that the host OS cannot bypass. For B200 clusters, the eSwitch and DOCA App Shield are the highest-assurance security controls available.
- UFM Cyber-AI is passive — it detects anomalies but does not remediate. Wire it into your SIEM. Define response runbooks for each alert category.
Lab 15 gives you hands-on experience reproducing an RKEY exposure attack, identifying both misconfigurations that enabled it, and applying GID filtering and RKEY rotation to harden the environment.