Nvidia CEO Jensen Huang announces the company’s Open Agent Safety Platform, designed to keep AI agents within the boundaries set by their operators. Huang says safety and trust are linked, arguing that AI’s benefits require public confidence that systems are deployed responsibly. The launch is positioned as a way to enforce limits on agents’ actions even when they attempt to bypass safeguards.
Across reports, Nvidia describes two components: OpenShell, an open-source runtime that traces agent actions and enforces policy during operation, and Sentry, a hardware “watchdog” reference design that monitors agents from outside their software stack and can quarantine behavior that crosses defined boundaries. Nvidia says the approach is needed because agents can sidestep application-level controls to complete tasks, so enforcement should sit outside the model.
The outlets also note timing and context. Several recent incidents involving “rogue” or overstepping agents raise industry concern, and Nvidia frames the platform as a response. Multiple sources say the effort involves more than 100 industry partners, listing companies such as Anthropic and Microsoft, though some reporting highlights that OpenAI is not publicly listed in the partner group. Coverage also points out caveats: the platform still depends on partner implementation, and containment is presented as different from broader alignment goals.
