NVIDIA Unveils Platform to Rein In Rogue AI Agents

NVIDIA has introduced the Open Agent Safety Platform, a new system designed to add enforceable safety controls to autonomous AI agents throughout their testing and deployment. The move responds to incidents where AI agents have bypassed application-level restrictions while completing tasks, prompting NVIDIA to build protections into the hardware and compute layers rather than relying solely on the AI model itself.

The platform has two main parts. OpenShell is an open-source runtime that tracks what an AI agent is doing and enforces rules about which systems, data and services it can access, using sandboxing and credential controls. It has already been integrated by Salesforce with Slack, letting staff review agent activity and approve or deny extra permissions. Sentry is an optional hardware watchdog built on NVIDIA's BlueField-4 chips that independently monitors agent behaviour from a separate, isolated environment and can quarantine an agent within milliseconds if it oversteps its boundaries.

More than 100 organisations, including Microsoft, CrowdStrike, Palo Alto Networks, SAP, Salesforce and ServiceNow, are already working with the platform, spanning sectors such as financial services, energy and robotics where AI agents may act with real-world consequences.