Today, NVIDIA announced the NVIDIA Open Agent Safety Platform, an open software platform and reference system design to strengthen AI security from agent testing to deployment, helping organizations secure and govern AI agents as they take on more autonomous work.

By Rob Thomas | Senior Vice President, IBM Software and Chief Commercial Officer

It's the right move at the right moment, and IBM is proud to support this effort, from agent identity and credentials to infrastructure, storage and hybrid cloud. Trust in AI agents requires more than model guardrails. It requires all of these capabilities working together, including with:

• IBM Agent Identity, currently in Public Preview, and HashiCorp Vault now integrate with NVIDIA OpenShell™, giving every agent a verified identity and only the access it needs.
• IBM Identity Protection helps clients find and monitor every agent, including the ones they didn't know they had.
• IBM Storage is integrating NVIDIA BlueField-4 directly at the hardware level in IBM Fusion, adding another layer of protection for the data agents access.
• Red Hat OpenShift provides the platform to run governed agents across the hybrid cloud, with support for NVIDIA BlueField DPUs. Red Hat is working with NVIDIA to integrate BlueField™ and OpenShell™ which offloads and isolates networking, security, and other infrastructure services from tenant workloads, provide zero-trust security, with real-time threat detection, isolation, and runtime protection directly at the infrastructure and kernel layers.

Why does this matter now?

AI agents are moving well beyond simple assistants. They can execute long-running tasks, access systems, use tools, invoke other agents, and make decisions on behalf of users and organizations.

That creates enormous opportunity. But it also introduces two fundamental questions for every enterprise deploying agents: How do you let an agent act on your behalf without giving up control? And how do you prove what it did?

The answer can't live in the prompt.

Agents can ignore instructions or be compromised through attacks like prompt injection. Real control has to be enforced by the environment around the agent from identity and credentials to runtime, network, infrastructure, and hardware.

That's the idea behind NVIDIA's approach. NVIDIA OpenShell places agents inside a policy-governed runtime, while NVIDIA Sentry uses NVIDIA DOCA on NVIDIA BlueField-4 to verify agent identity, enforce access policies for data, tools, APIs, and services, and provide attested telemetry to inspect agent behavior and detect deviations. BlueField-4 provides the hardware-isolated, host-independent foundation for these capabilities. Sentry can quarantine agents that attempt to move outside their boundaries in milliseconds, helping limit the spread of unauthorized activity. These controls remain operational even if the host or agent workload is compromised.

IBM adds critical enterprise capabilities across that stack. Agent Identity and HashiCorp Vault allow humans to delegate to agents, establish who an agent is, what authority it has and what it can access. Red Hat OpenShift provides the hybrid cloud foundation to run those agents consistently across environments. IBM Identity Protection gives security teams visibility into agent activity, including shadow agents. Together, these capabilities help make agent actions attributable, least-privileged and auditable.

And we're doing this in the open, so clients have choice and can build on the technology investments they already have. That's also why IBM joined NVIDIA as a founding member of the Open Secure AI Alliance, which is governed by the Linux Foundation.

The first chapter of AI was about what models could do. The next is about what enterprises can trust agents to do.