NVIDIA Open Agent Safety Platform targets agent trust, Cisco joins

For three years, the focus in AI has been on making models smarter — benchmarks, context windows, reasoning scores. The article argues that the more consequential question now is what agents should be permitted to do, because the more access and autonomy an agent gets, the more effective it becomes but the greater the risk. That tension, it says, makes trust rather than intelligence the real limit on how fast AI adoption can move.

The proposed answer is not a harness that constrains an agent from outside its context, but external control with full context of what agents are doing. NVIDIA OpenShell handles secure agent execution, while Sentry runs as an out-of-band watchdog on BlueField-4 DPUs, with attested telemetry and millisecond-scale quarantine. Cisco's contribution spans Cilium and Tetragon at the kernel, Hypershield across networks and clouds, AI Defense and DefenseClaw at the application layer, and Agentic Identity and Access Management for non-human identity and per-action authorization. Cisco Data Fabric powered by Splunk Platform is positioned as the system of record.

The reasoning leans on a past bet: security could no longer be something traffic passes through at the perimeter, so enforcement became hyper-distributed across the kernel, network, and cloud. The article contends agents need this even more, because they call tools they didn't write, deploy code they won't supervise, and delegate to agents in environments their operator doesn't own. The stated stakes are whether distributed enforcement — which produces distributed evidence — can converge into one system of record. That is likely to be the test of whether the platform delivers on its promise, and the article presents it as work the whole industry depends on rather than any one vendor's to solve.