NVIDIA Launches Open Source Security Platform to Rein In AI Agents

As concerns grow among the public and policymakers about the security risks of advanced AI systems, NVIDIA has released a new open source security platform aimed at AI agents. The Open Agent Safety Platform is described as a full stack system for governing and controlling AI agents across software, hardware, compute and robotics environments.

The platform includes two main tools. OpenShell allows organisations to test AI agents in sandbox environments before deployment, letting operators define which files, networks, tools, processes and credentials an agent can access, then verifying those restrictions hold up before the agent touches a live network. NVIDIA has also added security updates to its Bluefield 4 chip, which enables monitoring of AI agent behaviour and enforcement of security policies outside the main system.

More than 100 organisations, including Anthropic, Microsoft, Arm, Palantir and JPMorgan Chase, have committed to using the platform. NVIDIA CEO Jensen Huang described it as the start of an open ecosystem to build trust in agentic AI systems, arguing that safety and innovation go hand in hand.