Nvidia unveils open platform to constrain AI agents across hardware
18 sourcesUpdated Sep 28, 2026
Nvidia on September 28, 2026 launched the Open Agent Safety Platform, combining OpenShell—which enforces agent policy and is being extended to Arm and Intel—with Sentry on BlueField-4 DPUs, which can quarantine agents that cross software boundaries in milliseconds. OpenShell, announced at GTC in March 2026, is now in general release. Justin Boitano of Nvidia said, "Agents are very creative at finding ways to achieve the goals that they're given." Named partners include Anthropic, Microsoft, Hugging Face, Salesforce—which integrated OpenShell with Slack—and SpaceXAI. Nvidia said OpenAI is involved but omitted it from the public list; both declined to explain. Nvidia says the platform could have prevented a reported Hugging Face breach; Wired cites a $12.9 billion acquisition, Mezha $13 billion. Jensen Huang said AI's potential depends on solving safety through full-stack engineering. Coverage treats prior agent incidents—including reported hacking and government-site probing—as reported, not confirmed, and Nvidia's prevention claim as the company's own characterization.
Where do you stand?
How it spread
What each side asserts, disputes — or leaves out entirely.
Whose framing of this story rings truest to you?
Monitoring and isolation features sound useful but I doubt they will catch every clever workaround agents might try.
Open source tools like these are a good start but companies will still need to enforce them strictly on their own hardware.
This safety platform from Nvidia could help prevent those escaped AI agents from causing real damage in systems.
