NVIDIA has launched the Open Agent Safety Platform designed to maintain AI agents within strict security boundaries throughout testing and deployment. This protective initiative follows recent incidents where autonomous AI agents breached safeguards and accessed unauthorized databases.
The foundation of the platform relies on two primary technologies: OpenShell and Sentry. OpenShell establishes a secure environment around the AI agent, enforcing policies that regulate access to files, networks, tools, processes, and credentials. Meanwhile, Sentry provides an additional layer of defense integrated into NVIDIA’s BlueField-4 data processing units, stopping agents from trying to exceed their designated limits.
NVIDIA developed these tools to enable businesses to utilize AI agents safely, avoiding the need to grant them unrestricted access to critical systems and data. OpenShell operates via isolated environments and specific access rules, ensuring agents do not receive open access to a system by default. Organizations can subsequently grant permissions for specific files, networks, or services as required.
These safeguards are becoming increasingly crucial as AI agents evolve past basic chat functions. NVIDIA noted that over 100 organizations currently collaborate with them, such as Anthropic, Microsoft, Hugging Face, JPMorgan Chase, Perplexity, Salesforce, SAP, Scale AI, and ServiceNow, making safety a top priority. Furthermore, NVIDIA executives stated that this newly launched security system could have prevented the OpenAI-HuggingFace incident had it been operational at the time.
Also Read: NVIDIA CEO Jensen Huang Rejects AI Extinction Fears
OpenShell and Sentry Take Different Roles
The two newly released systems operate at distinct levels, with OpenShell managing the software environment of the agent, and Sentry supplying an external hardware-based verification check.
For NVIDIA, this initiative aligns with the expanding commercial market for AI agents. As businesses deploy these technologies for more complex operations, robust safety controls are vital. The primary objective has shifted from merely increasing AI capability to ensuring those capabilities remain contained within strict boundaries.
