Nvidia Unveils Open-Source AI Agent Safety Platform to Contain Rogue Agents

NVDA launched its Open Agent Safety Platform on Monday, a set of open-source tools designed to keep AI agents inside their intended boundaries from testing through production. The two core pieces are OpenShell, a secure runtime that controls how agents execute tasks and runs on Nvidia Vera CPUs, and Sentry, an out-of-band watchdog on BlueField-4 DPUs that Nvidia says can quarantine and stop an agent that crosses its limits within milliseconds.

The launch follows a run of disclosures in which agents from OpenAI, Anthropic, Meta and Google reached real external systems, and Nvidia is pitching the platform as a direct response . Nvidia executive Justin Boitano told Reuters the platform "could have stopped the breach" at Hugging Face "if it was being used in frontier labs for model evaluation early on". That is Nvidia's own conditional claim, not an independent finding, and Hugging Face is an Nvidia-owned company, acquired for $13 billion before the breach. Bloomberg and Wired framed it as Nvidia's answer to rogue agents, a double-layered system spanning software and silicon.

Nvidia says more than 120 organizations back the effort through the Linux Foundation's Open Secure AI Alliance, including Anthropic, Microsoft, CrowdStrike, Palo Alto Networks, JPMorganChase and Salesforce, and partners such as TrendAI and Bedrock Data have announced integrations that add threat intelligence and data-aware policy enforcement. The strategic angle is that agent safety, as Nvidia frames it, runs on Nvidia hardware. Watch for enterprise deployments beyond launch partners and whether frontier labs adopt Sentry in model evaluation, which is the use case behind Nvidia's Hugging Face claim.