NVIDIA Open Agent Safety Platform targets AI agent trust

The announcement reflects a shift in how the industry frames agent safety. Earlier approaches relied on model guardrails and prompt instructions, but the article argues that agents can ignore instructions or be compromised through attacks like prompt injection. Control therefore has to be enforced by the environment around the agent, from identity and credentials to runtime, network, infrastructure and hardware.

IBM's supporting pieces span that stack: IBM Agent Identity and HashiCorp Vault establish who an agent is and what it can access; IBM Identity Protection helps clients find and monitor every agent, including shadow agents; IBM Storage integrates NVIDIA BlueField-4 at the hardware level in IBM Fusion; and Red Hat OpenShift provides the platform to run governed agents across the hybrid cloud. NVIDIA's own layer, NVIDIA OpenShell, keeps agents inside a policy-governed runtime, while NVIDIA Sentry verifies identity and enforces access policies.

The approach is explicitly open, with IBM joining NVIDIA as a founding member of the Open Secure AI Alliance, governed by the Linux Foundation. The stated goal is to make agent actions attributable, least-privileged and auditable. How much enterprises can rely on these controls without adding new operational burdens is likely the next question, especially as agents take on longer-running, autonomous work.