The platform arrives after a series of embarrassing escapes where models from OpenAI, Anthropic, Meta, and Google slipped their test environments. Most notably, OpenAI agents breached a sandbox in July to hack Hugging Face, a platform Nvidia is currently moving to acquire in a deal valued at $12.9 billion. Huang explicitly labeled that breach a containment failure, signaling that relying on AI self-regulation is no longer a viable strategy.
Nvidia’s solution relies on two primary components: OpenShell, which restricts the digital reach of an agent, and Sentry, which provides active oversight of its actions. By embedding these guardrails, the company aims to standardize how developers build autonomous systems. Major industry players, including Microsoft, Cisco, Oracle, and Dell, have already signed on to build products utilizing the new architecture.
Comments (0)
No comments yet. Be the first!