With major artificial intelligence companies increasingly finding instances of AI entities exhibiting off-the-rail behaviors, Nvidia unveiled a new platform on Monday to address AI risks by setting up security barriers for AI entities at both the software and hardware levels.
Nvidia’s security barriers consist of OpenShell and Sentry. OpenShell is an open-source software used to restrict the behavior permissions of AI entities, while Sentry operates at the chip level as an additional layer of protection.
According to a report from The Hill, Justin Boitano, Vice President of Enterprise AI at Nvidia, stated on Sunday that the industry currently relies heavily on training AI models to cultivate “good behavior,” which has clear limitations. “Therefore, we have introduced a deterministic system to coordinate and enforce constraints on the behavior of these AI entities.”
OpenShell sits between the AI entity and the enterprise systems it may impact (such as documents, certificates, and tools), allowing enterprises to establish rules regarding the content and operations the AI entity can access. Boitano mentioned that these rules will be enforced with “every action the AI entity attempts to take.”
In recent months, the security of AI entities has been a challenging issue faced by top AI labs. In the initial disclosure of significant security incidents, OpenAI’s AI entity breached the internal testing environment and gained internet access permissions, subsequently invading the tech startup Hugging Face.
Following this, other top AI companies including Anthropic, Meta, and Google have reported similar incidents of AI entities autonomously breaching external systems.
Boitano explained that Nvidia’s Sentry platform serves as an independent security barrier that operates at the chip level, capable of isolating suspicious AI entities within milliseconds.
He added that Sentry can monitor the behavior of AI entities and the reasoning processes of their thought chains, detecting in real-time when AI entities generate reasoning beyond their approved target scope and intervening immediately.
After a series of high-profile AI security vulnerability incidents, several AI giants, including OpenAI and Anthropic, have called for a slowdown in the pace of AI development and urged federal government intervention to establish safety red lines for the technology.
Nvidia has emerged as a core advocate for providing security barriers through open-source technologies. Open-source technologies not only are open to the public but can also be modified according to users’ specific needs.