Lesson 3 — Check Point AI Factory Firewall on BlueField-3

You know the DPU (Lesson 1) and why security belongs on it (Lesson 2). Now the product. Check Point AI Factory Firewall — previously AI Cloud Protect — embeds Check Point's firewall and threat prevention directly into NVIDIA BlueField-3 DPUs via DOCA.^press

The one-sentence mental model

Check Point AI Factory Firewall turns each GPU server's BlueField-3 DPU into an inline firewall + host sensor — securing AI infrastructure at the infrastructure level, with zero impact on GPU/CPU.^brief

1 · The two layers

This is the answer you should be able to give in your sleep. AI Factory Firewall protects on two levels, each using something you already learned:^brief

Layer 1 — Network-level (inline, on the DPU data path — Lesson 1) next-gen firewall identity-based access IPS virtual patching anti-malware

＋

Layer 2 — Host-level (via DOCA Argus direct memory access — Lesson 2) sees all host processes stops local code execution blocks privilege escalation catches malicious LLM processes

Layer 1 watches the traffic (N-S and E-W between AI nodes); Layer 2 watches what's running on the node. Together: the wire and the host, from one isolated DPU.

2 · In the AI factory

Zoom out to the data center. Every GPU node carries a BlueField-3; AI Factory Firewall on each one becomes a distributed security layer wrapping the whole factory — including the lateral (east-west) traffic that traditional perimeter firewalls never see.

Outside / other tenants north-south traffic

▲ ▼

GPU node 1 BlueField-3 + AIFF
GPUs

GPU node 2 BlueField-3 + AIFF
GPUs

GPU node N BlueField-3 + AIFF
GPUs

◄ ─ ─ ─ east-west traffic between nodes — each DPU inspects it ─ ─ ─ ►

Security moves to every node instead of a chokepoint at the edge — so a compromised node can't freely reach its neighbors.

3 · The threats it's built to stop

AI factories face threats a generic firewall wasn't designed for. Name these — they're the "why buy it":^brief

Threat	Which layer answers it
Data poisoning (tainting training data)	Network: access control + inspection of data flows
Model theft / exfiltration	Network: block unauthorized egress; identity control
Malicious process in a downloaded LLM	Host: DOCA Argus process visibility
Local code exec / privilege escalation	Host: Argus runtime detection
Lateral movement between GPU nodes	Network: east-west inspection at each DPU

4 · How it's deployed & managed

Security runs on the DPU, isolated from host and AI workloads → no GPU/CPU consumption. Validated on NVIDIA RTX PRO Servers; packaged as an NVIDIA NGC container (nvidia/doca-checkpoint/aicp).^blog
Multi-tenanted management with predefined security policies designed for AI; standard orchestration for rollout across many nodes.^brief
Status: GA for on-prem enterprise (rebranded from AI Cloud Protect, Oct 2025). Roadmap: also runs on BlueField-4 (~6× compute, ~2× throughput).

Read the real docs — Check Point & SecureKnowledge

For deployment specifics use the AI Factory Firewall Getting Started Guide and search Check Point SecureKnowledge for "AI Factory Firewall" / "AI Cloud Protect". (This lesson cites public sources; SecureKnowledge carries the version-specific admin detail — always confirm against the current sk for your release.)

Check yourself

AI Factory Firewall's host-level layer is powered by…

Which is a Layer-1 (network) capability of AI Factory Firewall?

What is the headline benefit of running on the DPU?

Primary source — read this next

Check Point's AI Factory Firewall Solution Brief (the two-layer story) and the Getting Started Guide. Then browse SecureKnowledge for your release's admin articles.