NVIDIA has announced the NVIDIA Open Agent Safety Platform, an open software platform and reference system designed to help organisations test, secure and govern AI agents from development through deployment.
The platform is aimed at addressing security challenges emerging as AI agents move beyond conversational assistance to performing longer-running tasks, accessing enterprise systems, using tools and APIs, interacting with other agents and making decisions on behalf of users.
NVIDIA’s approach places AI agents within a policy-governed runtime environment, with controls extending beyond model-level guardrails to identity, credentials, runtime security, networking, infrastructure and hardware.
At the centre of the platform is NVIDIA OpenShell, which provides a controlled runtime for agents. NVIDIA Sentry, meanwhile, uses NVIDIA DOCA on NVIDIA BlueField-4 to verify agent identity, enforce access policies and provide attested telemetry for monitoring agent behaviour.
According to NVIDIA, BlueField-4 provides a hardware-isolated and host-independent foundation for these security capabilities. Sentry can detect deviations from established policies and quarantine agents that attempt to operate outside their authorised boundaries, including when the host or agent workload has been compromised.
IBM is supporting the initiative across several layers of the enterprise technology stack. IBM Agent Identity, currently in public preview, and HashiCorp Vault integrate with NVIDIA OpenShell to provide agents with verified identities and controlled access. IBM Identity Protection is designed to help organisations discover and monitor deployed agents, including previously unidentified agents.
IBM Storage is also integrating NVIDIA BlueField-4 at the hardware level in IBM Fusion, adding protection for data accessed by AI agents. Red Hat OpenShift provides a hybrid-cloud platform for running governed agents and supports NVIDIA BlueField DPUs, while Red Hat and NVIDIA are working to integrate BlueField and OpenShell to offload and isolate networking, security and infrastructure services from tenant workloads.
The initiative reflects the growing need for enterprise AI security architectures that can maintain control over autonomous systems even when agents interact with sensitive data, tools and infrastructure.
As agentic AI adoption expands, the platform seeks to provide organisations with mechanisms to establish who an agent is, what it can access, what it does and how its actions can be monitored, extending AI governance from model behaviour to the broader infrastructure supporting autonomous agents.
