Now loading...
NVIDIA has unveiled the NVIDIA Open Agent Safety Platform, an innovative open software platform and reference system designed to enhance AI security, from testing to deployment. This platform offers comprehensive governance and control over the entire stack of software, hardware, computing, and robotics systems that operate AI agents. The initiative is a response to recent security breaches highlighting the urgent need for organizations to have customizable tools that reinforce control over long-running AI agents, which have been shown to bypass security measures at the application level.
Jensen Huang, the founder and CEO of NVIDIA, emphasized that realizing the societal potential of AI hinges on addressing safety challenges. He stated that accelerating advancements in AI safety requires extensive engineering across the full stack. The Open Agent Safety Platform aims to unite industry leaders, researchers, and public-sector organizations to cultivate best practices, establish aligned evaluation methods, and promote international collaboration to enhance global AI safety standards.
The NVIDIA Open Agent Safety Platform is structured to provide governance and control over the software running the agents, as well as the hardware and computing layers that support these operations. Organizations can customize the deployment of platform elements to suit their specific needs. Among its key components is the NVIDIA OpenShell, a secure runtime software that creates enforceable boundaries for agents operating on CPUs. OpenShell can be extended to interact with third-party platforms, including those from Arm and Intel, enabling enterprises to safeguard agent operations across diverse systems.
The platform also includes the NVIDIA Sentry, a supplementary monitoring system that utilizes NVIDIA BlueField-4 data processing units to oversee agent behavior in real-time. Sentry’s robust design ensures that if an agent attempts to operate outside of its designated parameters, it can be swiftly quarantined and halted. This monitoring is critical for ensuring compliance with security standards, as it operates independently and in real-time, providing an additional layer of protection against potential breaches.
Industry giants from various sectors are rallying around NVIDIA to reinforce AI security. Notable collaborators include Anthropic, Microsoft, Dell Technologies, and SpaceXAI, among others. These partnerships are focused on developing mechanisms that provide greater oversight and verification of AI agent activities, especially in sensitive environments.
Anthropic’s collaboration aims to improve security by executing agent operations on distinct servers while maintaining strict control over their data and tasks. Similarly, SpaceXAI is employing the NVIDIA Open Agent Safety Platform for its coding models, emphasizing the necessity of establishing firm safety boundaries for AI agents.
In addition, Scale AI is working with NVIDIA to integrate these technologies into its agentic infrastructure portfolio, thereby ensuring that the systems used for critical applications adhere to stringent security and operational guidelines. Salesforce has also taken steps to connect OpenShell with Slack, enabling teams to monitor agent actions and manage permissions more effectively, thereby promoting enhanced visibility and control.
