IBM is supporting NVIDIA’s newly announced Open Agent Safety Platform, an open software platform and reference system designed to strengthen security and governance for increasingly autonomous AI agents from testing through deployment.
The initiative combines NVIDIA’s agent security architecture with IBM capabilities spanning agent identity, credential management, infrastructure, storage, and hybrid cloud, reflecting a broader push to secure AI agents not just at the model level but across the enterprise systems they can access and control. Building Trust Into the Next Ge…
AI agents are evolving beyond conventional assistants that respond to individual prompts. They can execute long-running tasks, access enterprise systems, use external tools, call other agents, and make decisions on behalf of users and organizations.
That shift creates new security challenges because enterprises need to determine what each agent is authorized to do, limit its access to sensitive systems and data, and maintain an auditable record of its actions.
IBM argues that those controls cannot be handled solely through prompts or model-level guardrails. Agents can ignore instructions or be compromised through techniques such as prompt injection, making controls across identity, credentials, runtime environments, networking, infrastructure, and hardware increasingly important. Building Trust Into the Next Ge…
At the center of NVIDIA’s approach is NVIDIA OpenShell, which places agents inside a policy-governed runtime.
NVIDIA Sentry uses NVIDIA DOCA running on NVIDIA BlueField-4 to verify agent identities, enforce access policies covering data, tools, APIs, and services, and provide attested telemetry that can be used to analyze agent behavior and identify deviations.
BlueField-4 provides a hardware-isolated foundation that operates independently of the host system.
According to IBM, Sentry can quarantine agents attempting to operate outside their established boundaries in milliseconds, helping reduce the spread of unauthorized activity even if the underlying host or agent workload has been compromised.
IBM is integrating several of its security and infrastructure products with that architecture.
IBM Agent Identity, which is currently in public preview, and HashiCorp Vault are integrating with NVIDIA OpenShell to provide AI agents with verified identities and credentials limited to the access required for their assigned tasks.
The combination is intended to help organizations establish who or what an agent is, determine the authority delegated to it, and specify which resources it can access.
IBM Identity Protection provides another layer by helping security teams discover and monitor AI agents operating across an organization, including potentially unmanaged or previously unidentified “shadow agents.”
This visibility is intended to give organizations a more complete inventory of autonomous systems interacting with corporate resources.
IBM is also extending security into its data infrastructure.
IBM Storage is integrating NVIDIA BlueField-4 at the hardware level into IBM Fusion, adding additional protection around the data accessed by AI agents.
For hybrid cloud deployments, Red Hat OpenShift provides an environment for operating governed agents across different infrastructure environments with support for NVIDIA BlueField data processing units.
Red Hat is also working with NVIDIA to integrate BlueField and OpenShell.
The architecture is designed to offload and isolate networking, security, and other infrastructure services from tenant workloads while providing zero-trust controls, real-time threat detection, isolation, and runtime protections at the infrastructure and kernel layers.
Together, IBM said these technologies can help enterprises make AI agent activity attributable, least-privileged, and auditable rather than relying entirely on behavioral instructions delivered to the AI model.
The approach is particularly relevant as organizations move toward agentic AI systems capable of interacting with sensitive databases, enterprise applications, APIs, cloud infrastructure, and other autonomous agents without continuous human supervision.
IBM and NVIDIA are also supporting an open approach to AI security.
IBM has joined NVIDIA as a founding member of the Open Secure AI Alliance, which is governed by the Linux Foundation.
The open model is intended to allow enterprises to implement security controls while continuing to use existing technology investments rather than becoming dependent on a single closed security architecture.
For IBM, the initiative extends its broader enterprise AI strategy into the security and governance layer required to deploy autonomous AI systems at scale.
As agents gain greater authority over business processes and enterprise systems, IBM sees identity, infrastructure-level isolation, access controls, and auditing becoming central to determining which autonomous tasks organizations are willing to delegate.
KEY QUOTES:
“Trust in AI agents requires more than model guardrails. It requires all of these capabilities working together.”
“The first chapter of AI was about what models could do. The next is about what enterprises can trust agents to do.”
Rob Thomas, Senior Vice President, IBM Software And Chief Commercial Officer
