Monday, November 23, 2026

NVIDIA Unveils Security Platform to Prevent AI Agent 'Jailbreaks'

Input
2026-09-29 11:06:16
Updated
2026-09-29 11:06:16
Reuters-Yonhap News

[Financial News] NVIDIA has unveiled a new software platform designed to prevent AI agents from breaking out of control and causing security incidents.
On the 28th local time, NVIDIA CEO Jensen Huang explained in an interview with CNBC, "We can't allow AI agents to roam unauthorized through internal systems, so we need to find a way to isolate them. The platform unveiled this time serves as a kind of 'agent-only browser' that limits agents to accessing only the resources they need to perform their tasks."
NVIDIA's NVIDIA Open Agent Safety Platform was developed in response to recent incidents involving major AI companies, including OpenAI, Anthropic, Meta, and Google, in which agents escaped sandboxes and attempted to infiltrate third-party systems.
An NVIDIA official referred to an incident in July in which an OpenAI model infiltrated Hugging Face and said the incident could have been prevented if the new platform had been in place. At the time, the OpenAI model escaped its sandbox and attacked the infrastructure of the open-source platform Hugging Face, with more than 17,000 agent attacks reported over several weeks.
Justin Boitano, NVIDIA's vice president of enterprise AI, emphasized that the platform is an engineering solution to address safety concerns, saying, "Recent incidents have revealed that model-level guardrails alone cannot fully control an agent's access permissions and actions."
The platform includes NVIDIA OpenShell, which operates on central processing units (CPUs) and sets limits on agents' permissions, as well as Sentry, which operates on network chips rather than CPUs or graphics processing units (GPUs) and monitors agents' movements in real time.
NVIDIA plans to release some of the platform's software as open source and offer it in the form of a reference design, enabling partners to build their own security products. Microsoft, Cisco, Oracle, Dell, HPE, Lenovo, Intel, and Arm are currently participating as major partners, while NVIDIA is also collaborating with Anthropic to integrate OpenShell.
As warnings about AI models becoming uncontrollable have recently mounted among the heads of major technology companies, including Anthropic, OpenAI, and SpaceX, CEO Jensen Huang has proposed a direct technological response. Huang added, "There can be no successful AI industry without the belief that AI is built and deployed safely."
[email protected] Yoon Jae-jun Reporter