Red Hat and NVIDIA move AI agent security into BlueField hardware on OpenShift

Two failure modes for autonomous agents

In a blog post dated October 21, 2026, Red Hat product manager Gurpreet Singh argues that autonomous agents are powerful because they act, but that this also makes them a new category of infrastructure risk. When instructions are ambiguous or a task takes an unexpected turn, an agent makes judgment calls, and according to the post, "You can't expect an agent to fully police its own behavior."

The post names two failure modes that emerge as enterprises scale agentic AI: agents that overstep their operational bounds, and agents that are weaponized by external attackers. The NVIDIA OpenShell Reference Architecture addresses both, Red Hat says, by moving away from relying solely on model guardrails. Instead it sets agent runtime boundaries with OpenShell and adds hardware-isolated monitoring and policy enforcement with Sentry on BlueField.

How Sentry and DOCA Argus work

The architecture rests on two principles: governance, meaning agents only perform actions they are explicitly authorized to do, and security, meaning compromised agents are detected and contained quickly. It includes NVIDIA Sentry, a unified security software framework built on NVIDIA DOCA that runs on BlueField for out-of-band monitoring and policy enforcement.

Sentry uses DOCA Argus for security monitoring and runtime threat detection in containerized environments, including Red Hat OpenShift. According to the post, Argus works from the infrastructure layer and performs live machine introspection. It inspects selected host-memory structures from BlueField's separate trust domain and delivers runtime telemetry and threat alerts without a monitoring agent on the host.

Argus observes process, file and network activity. If it detects suspicious activity or a policy violation, it generates an alert that an integrated security system can use to trigger enforcement. To identify suspicious runtime activity, it uses process information, network connections and the hashes of running executables.

The OpenShift foundation

The setup builds on generally available support for NVIDIA BlueField-3 DPUs in Red Hat OpenShift 4.22. OpenShift manages the environment with a dual-cluster topology through the DOCA Platform Framework (DPF). The tenant cluster runs on the server's main CPU cores and hosts the AI agents. The infrastructure cluster uses BlueField DPUs as worker nodes to run networking and security services on the Arm processor.

This physical separation creates a separate infrastructure trust domain. In Zero Trust mode, BlueField restricts host access to its management interfaces, which protects DPU-resident controls and telemetry even if the host is compromised, according to Red Hat.

For containment, the DPF operator manages DPU policies natively. An integrated security system can then have BlueField isolate the affected workload's Virtual Function (VF) at the hardware level. Red Hat says Sentry can quarantine a rogue agent in milliseconds, without consuming CPU cycles or alerting the attacker.

What it means for enterprise teams

Red Hat's case for hardware enforcement is that offloading the security and networking stack to the BlueField DPU reduces CPU overhead and frees host CPU resources for revenue-generating AI workloads. Uniform policy enforcement across every node, it says, helps limit the spread of a compromise beyond the affected workload. The post concludes that bolting security onto the host operating system is no longer sufficient when deploying agents at scale.

The Red Hat AI Factory with NVIDIA combines out-of-band monitoring with DOCA Argus on OpenShift, DPF-managed policies to isolate workloads, and OpenShell as a secure agent runtime. According to the post, this lets security teams monitor and respond when software inside the agent workload cannot be trusted. The post does not give pricing or performance benchmarks beyond the millisecond quarantine claim.