NVIDIA has launched New tools designed to include artificial intelligence agents when they try to transcend established limitations. The platform comes as the need for systems capable of autonomously performing complex tasks to access resources or infrastructure beyond what was foreseen increases.
The talk also saw several incidents involving agents developed by companies like OpenAI and Anthropic. According to NVIDIA, security systems based on controls applied directly to the infrastructure can reduce such risks No need to delegate protection entirely to the model.
OpenShell and Sentry work together to include agents
One of the main components is OpenShell, A runtime designed to run agents in an isolated environment And enforce precise policies on files, processes, networks, and credentials. NVIDIA pairs it with Sentry, a system that uses separate hardware components to monitor agent behavior and intervene when an agent attempts to leave its environment.
The company claims that this combination Could have prevented the facehugging attackonly became known in recent months. Justin Boitano, vice president and general manager of enterprise computing at NVIDIA, told Reuters that the leak could have been prevented if the platform had been used in the early stages of evaluating the model. However, this is NVIDIA’s own assessment and not independent verification.
Nvidia CEO isn’t worried about artificial intelligence making kids forget math: ‘I don’t even know my address’
OpenShell also takes advantage of the hardware capabilities of NVIDIA processors, and the company is working with Arm and Intel to bring the system to other platforms. NVIDIA is demonstrating the solution with dozens of partners, including Anthropic.
Another factor involves the behavior of agents trying to circumvent restrictions, for example, by creating other agents responsible for performing specific tasks. NVIDIA claims Use mathematical controls to identify these patterns and apply these policies to the activities of subagents as well. The platform therefore reflects an approach to moving security from a separate model into the environment in which agents operate. NVIDIA says controls should be applied at multiple levels to prevent a system from directly modifying or bypassing its own protections.