Nvidia unveiled a new security platform designed to prevent artificial intelligence agents from breaking free of controls, malfunctioning or taking dangerous actions. The announcement follows a string of disclosures by major AI companies that their own models had escaped oversight and broken into other organizations' systems.
Nvidia CEO Jensen Huang introduced the Open Agent Safety Platform during an appearance on CNBC's "Squawk Box" on the 28th, describing it as "a browser for agents." Huang said companies cannot let agents roam freely across their systems and must find ways to confine them to specific zones.
According to Nvidia, the Open Agent Safety Platform includes open-source software that sets the boundaries of an agent's behavior. Justin Boitano, Nvidia's vice president for AI, said software called OpenShell allows developers to formally verify that an agent has enough authority to carry out its assigned task and no more. Because it is open source, the technology can be extended to run on rival computing platforms, including those from ARM and Intel.
The platform also includes Sentry, a separate security layer embedded in the chip that continuously monitors AI agent activity and can intervene immediately if an agent strays from its objective. Boitano said suspicious material can be isolated within milliseconds, or thousandths of a second. He said OpenShell manages agent behavior while Sentry independently watches for and blocks suspicious activity, adding that more than 100 organizations — including Microsoft, Perplexity, Accenture and JPMorgan Chase — have been using the platform since launch.
Nvidia executives said in a press briefing that the new system could have prevented the recent Hugging Face hacking incident involving a swarm of OpenAI agents. Boitano said the breach could have been stopped if the new security platform had been used from the outset at the frontier labs that evaluate models. The Hugging Face hack was a major intrusion that raised significant safety concerns about AI, and was followed by similar malicious activity involving OpenAI, including a breach of an Australian health ministry website. After OpenAI, companies including Anthropic and Meta also disclosed that their AI systems had hacked other organizations.
The incidents have fueled debate over the safety of advanced AI systems, including concerns that self-improving AI models could slip beyond human control.
Huang stressed that a successful AI industry is impossible unless people are confident AI has been built and deployed safely. Nvidia is working with Anthropic — a partner of Cisco, Microsoft and Oracle — to integrate cloud management agents with OpenShell.
Separately, Nvidia said the same day that it had authorized an additional $150 billion for its share buyback program, raising the total to $235 billion, the largest such program in history. The company plans to complete all remaining buybacks by fiscal 2028 as surging demand for AI training and inference accelerates its cash generation.
