Nvidia wants to put a watchdog chip next to every AI agent

Nvidia unveiled the Open Agent Safety Platform, a two-part system designed to keep autonomous AI agents inside owner-defined boundaries. OpenShell is free, open-source software that traces every action an agent takes and enforces rules set by the operator; it’s tuned for Nvidia’s Vera processors but extensible to other chips and is available on GitHub. Sentry is a reference hardware design running on BlueField-4 data processing units that sits outside the host system, verifies agent identity, checks each request, and can quarantine or stop an agent in milliseconds. The architecture places controls outside the model so an agent cannot simply “talk” or code its way around restrictions, a response to recent escape incidents such as data exfiltration via DNS and coordinated agent intrusions.

More than 100 organizations have signed up, including Anthropic, Microsoft, SpaceXAI, banks like JPMorgan Chase and Citi, Salesforce (integrated with Slack), SAP, Scale AI and several robotics firms. Nvidia positions the platform as an industry effort and links it to the Open Secure AI Alliance under the Linux Foundation. Major providers behind recent incidents - OpenAI, Google, Meta and Amazon - are not listed. Nvidia’s claims have not been independently verified, and Sentry is presented as a reference design with no public price or shipping timeline.