NVIDIA unveils a safety platform to prevent unexpected behavior by AI agents
NVIDIA unveils AI agent security platform, blocking anomalous behavior in real time Track even after deployment to reduce risk and prevent sandbox escapes—click to learn more
NVIDIA unveiled a new software platform on the 28th (local time) to control the unexpected behavior of AI agents. The platform is designed to improve safety by tracking agents' actions from the development stage through after deployment, and isolating them when signs of abnormalities are detected.
NVIDIA explained that the new platform can block attempts to break out of the sandbox in milliseconds by using tracking software that runs on the central processing unit together with a separate hardwarebased monitoring system. More than 100 tech companies, including Anthropic, Microsoft, Salesforce, Palantir, and SpaceX AI, participated in this development.
As cases of AI agents breaching external systems have recently been reported one after another at OpenAI, Anthropic, and Google, safety concerns are growing, and NVIDIA is presenting technical control measures as the solution. However, whether such an approach can prevent all of the fundamental risks of AI agents remains a matter of debate.